The Quiet Revolution in Machine Vision

A state-of-the-art survey paper, including historical review, perspectives,

and future directions

Melvyn L Smith, Lyndon N Smith, Mark F Hansen

Centre for Machine Vision,

Bristol Robotics Laboratory, University of the West of England, Bristol, UK

Abstract
Over the past few years, what might not unreasonably be described as a true revolution has
taken place in the field of machine vision, radically altering the way many things had
previously been done and offering new and exciting opportunities for those able to quickly
embrace and master the new techniques. Rapid developments in machine learning, largely
enabled by faster GPU-equipped computing hardware, has facilitated an explosion of
machine vision applications into hitherto extremely challenging or, in many cases, previously
impossible to automate industrial tasks. Together with developments towards an internet of
things and the availability of big data, these form key components of what many consider to
be the fourth industrial revolution. This transformation has dramatically improved the
efficacy of some existing machine vision activities, such as in manufacturing (e.g. inspection
for quality control and quality assurance), security (e.g. facial biometrics) and in medicine
(e.g. detecting cancers), while in other cases has opened up completely new areas of use, such
as in agriculture and construction (as well as in the existing domains of manufacturing and
medicine). Here we will explore the history and nature of this change, what underlies it, what
enables it, and the impact it has had - the latter by reviewing several recent indicative
applications described in the research literature. We will also consider the continuing role that
traditional or classical machine vision might still play. Finally, the key future challenges and
developing opportunities in machine vision will also be discussed.

Keywords: machine vision; machine learning; deep learning; state-of-the-art

1. Introduction – the purpose of this paper and what it brings that is new to the reader
The appropriate utilisation of visual data is the enabling component in many automated tasks,
from industrial quality control to driverless cars. Here we consider the recent history of
machine vision research, the huge changes that have, and are still taking place, and the
dramatic impact this is having, both in terms of methodology and new applications. We
introduce, define, and explain key terminology for those not completely familiar with the
discipline, and comment on the most relevant and impactful developments by pinpointing
ideas in the literature which have changed radically the approaches and techniques now being
explored. Using insightful examples taken from representative applications, we will explain
how these ideas are being applied now and consider the future directions.

2. What is machine vision?

The introduction of machine vision to industrial processes is often motivated by a desire to
reduce costs by increasing efficiency (and so productivity), reduce errors (and so improve
quality) or gather data. Equally importantly, it may also substitute for an absence of available
skilled labour or release workers from dangerous, demanding or fatiguing industrial
activities. In the past, the definition of the term has been somewhat unclear, however more
recently ‘machine vision’ has, largely de facto, come to be understood as the practical
realisation of image understanding, or more specifically computer vision techniques, to help
solve practical industrial problems that involve a significant visual component. The more
recent emphasis on the marriage between machine vision and machine learning, that has so
revolutionised the discipline, has been made possible by transformative developments in the
field of artificial intelligence (AI) and has served to move some machine vision capabilities
closer to that of human vision – a long unmet ambition that has existed since the early days of
computer vision as far back as the 1960s.

3. Machine learning and the artificial neural network (ANN)

To help appreciate why this transformation has come about, we first need to understand the
role of machine learning and have an appreciation of some of the overlapping terminology.
Deep learning is a particular kind of machine learning, which in turn is itself a subset of
artificial intelligence or AI (Figure 1). The term general AI, also referred to as strong AI,
captures the ambitious notion of a theoretical synthetic device able to learn, reason and
behave as humans do. This does not (perhaps yet) exist. A form of AI that does exist is that of
narrow AI. This includes software able to dramatically outperform humans, but only in very
limited, usually single, tasks. It is generally the case that while machine learning software can
automatically learn from historical data, and so improve with experience, and make informed
decisions using a range of (often statistical) techniques, without needing to be explicitly
manually programmed, the term deep learning is reserved for a special kind of machine
learning that allows computers to solve much more challenging and complex problems.

Deep
Learning

Artificial
Machine Intelligence
Learning

Figure 1 The relationship between the domains of AI, machine learning and deep
learning

In practice, deep learning is realised using artificial neural networks (ANNs) to simulate
human-like decision making. Inspired by, and loosely simulating the functioning of the
human cortex, ANNs comprise a matrix of connected layers of nodes (Figure 2). The input
layer is where data enter the network (for which the number of nodes is the same as the
dimension of input features). This is followed by one or more hidden layers that transform the
data flowing through the network towards the output layer (where the number of nodes is the
same as the number of classes to be classified), that produces the network predictions. A deep
network simply refers to an ANN with more than three hidden layers and in some cases can
have millions of nodes (loosely analogous to neurons in the human brain). We should note
here that in fact there are two different kinds of output prediction - either regression (a
continuous quantity), or, as mentioned above and which is more often the case in many
machine vision applications, a classification (a discrete class label or category).

Inputs Outputs

Input Layer Two Hidden Layers Output Layer

Figure 2 The connected node topology of an artificial neural network (ANN)

Connections between the layer nodes have weights and biases, known as the layer parameters
(Figure 3), that control how the layers together transform data between the input and output
of the network during a process known as feedforwarding. It is these parameters that
represent the network model, and we shall see later how a feedback learning process is used
to first establish these parameters during training. The development of ANNs has its origins
as far back as the 1950s and for a while fell out of favour. However, a series of breakthroughs
dating from the 1990s by Yann LeCun et. al. [1, 2], culminating in [3] and then later around
2010, most notably the work by Alex Krizhevsky [4] and others in image classification, led to
a particular form of ANN known as a deep convolutional neural network (CNN) becoming
established for many challenging computer vision tasks, particularly those involving
perception for object recognition and localisation in natural images.

Bias
Input 1 Weight 1
Output

Input 2
Weight 2

Figure 3 The layer node parameters (Output = Input 1 x Weight 1 + Input 2 x

Weight 2 + Bias)

3.1 The convolutional neural network (CNN) – a special kind of ANN for images
Figure 4 shows the principal components of a convolutional neural network (CNN). The main
difference between CNNs and regular ANNs, and the reason why they lend themselves so
well to image analysis, is the addition of a convolutional layer - an initial filtering stage,
implemented using conventional filter kernels that traverse the 2D image to detect local
spatial features, such as edges, corners, and other patterns, at different positions – somewhat
like processes in human vision. Outputs from the kernel are summed using a bias function to
form the input to a non-linear activation function layer, often a rectified linear unit (ReLU) (a
non-linear transformation that reduces negative input values to zero while positive values are
passed as output). This non-linear characteristic of the ReLU is also similar to the firing
behaviour of neurons in the human brain and so is suited to visual tasks (or more formally it
offers a reduced sensitivity to the vanishing gradient problem – an issue that impedes CNN
training). A data reduction stage (necessary as images may contain millions of pixels and so
vast amounts of data – some potentially redundant to the task) is then performed by a pooling
layer. This usually employs either average- or a max-pooling function to a defined
neighbourhood of pixel values, reducing the data dimensionality to leave only the most
significant pixels. Note here that the combined convolution and pooling layers may be
repeated, to detect a hierarchy of image features, before input to a fully connected
classification stage, giving a feature vector, and finally an output layer, for which a Softmax
activation function allows multinomial labelling (where the probability of all outputs sums to
1.0). Hence, the network output gives a probability distribution over the output labels
indicating which output case the network has most closely matched to the input. So, CNNs
are very well adapted to machine vision tasks, as they are specifically designed to process
large quantities of pixel data. The convolution and pooling layers act as feature extractors
from the input image, while the fully connected layers act as a classifier. In this way, CNNs
can usefully be divided into two parts: feature extraction and classification. Feature extraction
is achieved by the convolution plus ReLU together with pooling layers, and classification is
achieved by the fully connected layers.

Flower

Weed

Stone

Input Conv ReLU Pool Fully Connected Softmax Outputs

Feature Extraction Classification

Figure 4 The architecture of a convolutional neural network (CNN). The CNN can be
divided into two parts: feature extraction and linear classification

However, before all this can happen, the network must first be trained. Training enables the
network to output predictions by identifying patterns in a set of labelled training data, fed
through the network while the outputs are compared with the actual labels by an objective
loss function. During training the network’s parameters (the weight and bias of each neuron)
are tuned over a series of iterations or epochs, until the patterns identified by the network
result in good predictions for the training data. Thus, the parameters of the network (weights
and biases) are found during a supervised training phase by minimizing a loss function,
(calculated during a forward propagation), between prediction and ground truth labels at the
output layer. To help this process, regularisation constraints are used to update the network
weights and biases at each iteration (e.g. employing stochastic gradient descent – SGD) using
backpropagation until convergence. The concept of backpropagation in training the network
is very important and the key part of how an ANN is made to work. It is therefore worth
spending a little time underlining what is going on. During backpropagation, within each
layer, the node weights are adjusted in proportion to the amount of error they contributed to
the loss function (the difference between predictions and labels) during forward propagation.
The more a given node contributes to the error, the more it is adjusted. When the loss
function reaches a convergence state, the current weights and biases are preserved as the
well-trained model. It is in this manner that the weights and biases of the network are
established via the learning process using real data to form the deep learnt model. This
trained model may later be applied (a process known as inferencing) to unknown data (known
as generalisation) in a practical application to identify various features of interest in images.
CNNs have become the default choice for simultaneous detection of multi-class
categorisation problems, typical of many industrial visual tasks.

4. Classical machine vision

Since the 1980s, increasing computational capacity at a reducing cost, combined with
improvements in camera and lighting technologies, have made possible the practicable
deployment of industrial grade cameras and microprocessors together with new capabilities
in LED lighting (offering more control of the lighting intensity, colour or frequency and
differing projected patterns), to undertake a multitude of industrial vision tasks. Examples
include measuring the size and position of objects, assessing their surface quality in terms of
colour or texture, identifying defects or contamination, sorting, checking the integrity of
safety critical components and assemblies, or guiding industrial robots. In almost all cases,
these tasks have required strict environmental structuring, whereby, for example, any external
ambient lighting is eliminated and replaced, and the objects to be inspected are themselves
unchanging and in known fixed locations. In many production processes, where
manufactured objects are to be detected, measured, or inspected in some way, this
requirement for strict environmental structuring is relatively easily accommodated, as the
components being manufactured are already under control and most often in predictable fixed
positions at repeatable known times during the manufacturing cycle. For example, in the case
of objects on a production line conveyor or being handled by an industrial robot, there often
exists good opportunity to introduce cameras and lighting at suitable physical locations that
have a minimal impact on the production process. However, while this is fine in many
manufacturing production lines, there are also a wide range of other industrial activities
involving a visual sense where such environmental structuring is either problematic,
undesirable, or even completely impossible.

4.1 Why the need for structure in classical machine vision?

4.1.1 A rule driven vs a new data driven approach
For an automated system to be able to make sense of visual data, a model of the visual scene
together with a set of task rules and parameters is required. Conventionally, this model is
hand-crafted in the form of rules based on a priori information that is realised in code. In
other words, in a manufacturing scenario a CAD model of the object being detected,
measured, or inspected, is most likely readily available. In the case of defect detection, the
type and range of defects will also be known in advance of inspection system hard and
software design and installation. Also, given the nature of the manufacturing process, the
position and pose of the objects to be inspected, when viewed by a camera, can also be
established in advance, and are able to be readily fixed. Consider the example of the mass
production of kitchen and bathroom ceramic tiles [5], where an inspection stage is needed to
check conformity in terms of size and geometry and the presence of any unacceptable surface
anomalies. The latter may take the form of 2D colour or 3D moulding defects [6]. In this
case, a model describing the features of an ideal perfect tile can be obtained using existing
data (or alternatively using suitable good quality examples, sometimes called a ‘golden
template’ approach), and tolerances of acceptability applied to any variation in these features
captured during the production process. The introduction of structured lighting, along with
suitable camera and optics, can allow an ideal and repeatable view of the tiles in a fixed
position as they pass along a moving conveyor. All this available a-priori knowledge and
structuring reduces variation, offers predictability, and so greatly simplifies the machine
vision activity. Any ambient lighting, that may be subject to change, can be excluded, and
controlled artificial lighting substituted, allowing activities such as segmentation, i.e. the
isolation of the tile object and its features, to, for example, be performed using simple fixed
image intensity thresholds. Together, this structuring reduces the need to be able to
accommodate change and uncertainty within the inspection process. This means that the
machine vision software solution, that is the design of the algorithms realised in software, is
relatively straightforward and so easily hand-crafted.

Next consider an entirely different class of industrial machine vision problem, for example
that of identifying weeds within crops in a field as part of an agricultural farming application
[7]. This is typical of many machine vision tasks that involve natural objects, such as plants
(e.g. in farming, and in food and timber processing), minerals (e.g. polished marble and
granite), animals and animal by-products (e.g. meat and leather products) or humans (e.g. in
medical and security applications), as opposed to synthetic man-made objects – such as
manufactured metal and plastic parts and assemblies. In our example case of automated
weeding, the aim is to automatically provide a targeted precision dose of herbicide only to the
weeds, and not the crop or surrounding area [8]. Such a visually informed system might not
only reduce waste but could offer significant environmental benefits by minimising levels of
toxic chemical runoff and so ground and river pollution. The machine vision system
specification required here will need to be such that the device can both recognise the weeds,
or even perhaps a range of differing weed species (dock, ragwort, etc), and determine their
individual spatial locations. All will need to be completed in real-time in an outdoor
environment, as a tractor-mounted vision-equipped computer-controlled spray boom
traverses the crop. These uncontrollable variables, such as different invading species, the
amount of weed and crop growth, uneven ground, changing lighting and occlusion from the
crop itself, make the task inherently difficult as it includes a great deal of random variation.
Until very recently such tasks were considered entirely beyond the capability of a hand-
crafted rule-based model approach for any sort of practical application. Even so, such a
visually complex task is representative of many that are now being successfully addressed by
developments in machine learning techniques. More specifically, by deploying a form of
specialised artificial neural network, known as a convolutional neural network (CNN), it is
possible for the rules of such a complex machine vision model to be automatically learnt
from the data by the system itself, in a so-called data-driven approach.

5. Deep learning is changing the way machine vision is done

5.1 The conventional machine vision process
In conventional or classical machine vision, there are several well-established logical steps
necessary to realise almost any industrial machine vision application task. These can broadly
be categorised into the five core activities of: 1. image acquisition; 2. pre-processing; 3.
segmentation; 4. feature extraction; and 5. classification or interpretation (Figure 5). Image
acquisition will normally include aspects of environmental structuring, involving scene
constraints related to object positioning, camera and lens selection, and the design of a
suitable structured lighting arrangement. This is arguably by far the most impactful stage. It
needs to be very carefully considered as it can dramatically affect the complexity and so the
cost, efficiency, and reliability of the subsequent stages. The output following classification /
interpretation may take the form of an action, such as operating an actuator, guiding a robot,
altering a process variable (in the case of an adaptive closed-loop feedback system) or
generating a report (say in a quality or process control monitoring application). However,
most significantly in terms of the overall performance of the machine vision solution, are the
stages in between. This is where the creative thinking of the engineer developing a solution
largely takes place. Realising an optimal solution, often via a time-consuming trial-and-error
process to establish rules and tune parameters, depends strongly on the knowledge,
experience, and judgment of the system developer. Note also that when combined, these
stages produce a translation of data type, along with a filtering, to give a dramatic reduction
of data quantity. What starts as an optical image, is converted into a numerical data 2D array,
from which features and then scene descriptors are subsequently derived. So, while an
acquired image may be comprised of millions of bytes of data, say from a high-resolution
colour camera, the desired output from a notional machine vision system may take the form
of only a single binary bit, 1 or 0, indicating that, say, an object is either of acceptable quality
or not.

Optical 2D Data 2D Data Features Scene

Image Array Array Classification Descriptors
Image Feature
Pre-processing Segmentation or
Acquisition Extraction
Interpretation

Figure 5 The five steps of classical machine vision

5.2 The deep learning machine vision process

5.2.1 Deep learning is changing the role of the machine vision solution developer
The development of a machine learning, or more specifically a deep learning solution to a
machine vision task, is quite different from the processes involved for implementing the
conventional machine vision pipeline. In some ways the task is greatly simplified by
combining segmentation, feature extraction and classification stages, and it completely avoids
the need to design and hand-code these steps. Instead, the system developer is required to
possess a somewhat differing set of skills, including expertise in implementing deep learning
network architectures (e.g. selection of an appropriate network architecture, methods of data
augmentation, adjustment of network hyperparameters such as number of training epochs,
learning rate, L1 and L2 (regularization parameters), etc) [9]. As we have already seen, the
way the deep learning machine model is fashioned is via a process known as end-to-end
(usually supervised) learning, in which the ANN model is created by presenting the network
with labelled example image datasets (ground truth data). This labelling is used to identify
the object classes present in the image for training the network. As described above, the
labels represent the desired output for a given input and are used to create the loss function
that the network tries to minimise by adjusting the weights during backpropagation. Because
it is necessary to label the data, to explicitly tell the network what the data represent, this is
known as a supervised form of learning. The network learns the significant image features
needed to automatically map image inputs to target labels using a series of data
transformations. Then later in application (the network inferencing stage), the output of the
trained network takes the form of predictions, i.e. which of the possible outputs the network
thinks most closely match the given input. However, while this all sounds relatively
straightforward, there are a few downsides. Perhaps most significant of which is the amount
of example, manually labelled in the case of supervised learning, data (perhaps tens of
thousands of images) and time (often hours or days), needed to train the network. We will see
next that there are various ways to tackle these issues (along with a few others), however the
main takeaway here is that once trained, the inferencing (operational) stage is very much
quicker (perhaps taking only a matter of milliseconds).

6. A few notable key challenges with deep learning

Apart from the amount of time needed for training, there are one or two other challenging
issues regularly encountered when using deep learning to solve industrial problems that
deserve some special attention.
6.1 Limited training data: Often there may only be limited training data available for the
desired task - not enough to train a network from scratch. In such cases it may be possible to
deploy transfer learning with fine tuning as a feature extractor. Transfer learning is a machine
learning technique where a model (network) previously trained on one task is re-purposed on
a second (usually) related task. The advantage is that huge quantities of training data are not
needed. This can also help reduce hardware costs for the computationally expensive training
stage [10]. Here a pre-trained CNN (where the weights and biases are already established,
often by training on a large natural image dataset, for example the ImageNet dataset) is
subject to further supervised training for part, or all, of the network layers using the new
available data for the task under solution. Often, the last two layers in the original network
are replaced by two fully connected layers. The last layer is the output layer and is matched
with the number of classes in the dataset. Examples of differing pre-trained classifiers include
VGG-16 and VGG-19 networks [11], the ResNet50 network [12], Mobilenet [13], and
Xception [14]. Available data will usually be divided into training, validation, and test sets
for learning from examples, establishing the soundness of learning results, and evaluating the
generalisation ability of a developed algorithm on unseen data, respectively. Also, in the case
of limited data availability, cross validation methods (e.g. one-leave out, fivefold, or tenfold
validations) can be used.

6.2 The need to manually label training data: Most applications of CNNs involve
supervised training with large quantities of data – thousands of images. This can pose two
problems. Firstly, as mentioned above, is the limited availability of such large datasets for
training, and secondly there is the need to label (i.e. say what the image contains) the data
(usually a manual process). Both issues can to some extent be addressed using data
augmentation. This is where a limited labelled dataset is slightly altered using a range of
automated image transformations, such as rotations, reflections, height and width shifts,
zooming, horizontal-flipping, shear intensity changes, cropping, and adding noise, to
artificially create variation in the images and so a much larger training dataset. It is worth
noting that while such techniques can help improve training with a limited dataset, it is often
the case that use of transfer learning and data augmentation would not be expected to result in
as good a performance as training from scratch using a large dataset.

6.3 Overfitting the training data: Another well-known problem in machine learning is that of
overfitting. This is where the model fits the training data well but does not generalise to
unseen data (the useful part). Overfitting is one of the most frustrating problems in machine
learning projects and most practitioners experience it in their work. There are thankfully a
range of tools that can help, including just simplifying the network by reducing the number of
hidden layers, regularisation (adding a cost to the loss function for large weights) and the use
of dropout layers.
6.4 The mysterious black box: It is worth noting that it can be difficult to know what the
ANN is actually doing - for example which parts of the image are being used? Consider an
object recognition task. Could the network be using something unexpected, such as a
timestamp in the corner of the image and no part of the object itself? This can be of most
concern in medical applications, where even if performance is good, this lack of
accountability could have important legal consequences. Thankfully, a prediction process
debugging tool exists in the form of the Grad-CAM method. This gives a heatmap of class
activation that distinguishes the similarity of each image location with respect to a particular
class. More specifically, the heatmap of class activation is a feature map of the last
convolutional layer for a given input image, with the feature map channels weighted by a
class gradient calculated with regards to the feature map [15]. Figure 6 shows an example
heat map indicating the areas of a pig’s face used in a biometric recognition task [16]. This
shows in some detail which parts of the face the network was using for recognition and helps
us, to some extent, understand what the network was doing and the predictions it made.

Figure 6 Example heat map (right) indicating the areas of a pig’s face (left) used in a
biometric face recognition task where yellows and reds progressively show more
significant areas

7. Examples of machine vision tasks enabled by deep learning - a new class of machine
vision tasks
We next delve a little deeper into the details of how deep learning is currently being applied
to real machine vision industrial tasks via a range of seminal state-of-the-art real-world
prototypical example applications. We explore applications across differing sectors, consider
their context, including the role of big data [17] and the internet of things [18], the
motivations, and the potential advantages. We also review the common architectures in use
today; and look to the future by considering the most significant remaining problematic
issues still to be addressed.

7.1 Agriculture – pre-farm gate

As previously discussed, deep learning has improved performance in many existing industrial
tasks as well as allowing new areas of application. Perhaps none more so than in agriculture,
where, in so called ‘pre-farm gate’ applications, the imposition of any form of environmental
structuring is most often extremely difficult, and so in practice tends to be minimal. These
applications may include, crop assessment (detecting abiotic and biotic stress), pruning
(including weeding) and picking or harvesting (identifying ripeness and location).
Uncertainty and variation can exist both in terms of the natural objects being imaged and
their outdoor context.

Williams et al. [19] present the design and evaluation of a kiwifruit harvesting robot intended
to operate autonomously in pergola style orchards. The technology is typical of many agri-
tech solutions currently under development and the application highly topical as the work is
in part motivated within the context of a reduced availability of farm labour. A deep neural
network with stereo matching was used to detect and locate kiwifruit under real-world
conditions (with some added illumination). The machine supervised learning model utilised
here is typical of many similar applications and so worth exploring in a little detail. Semantic
segmentation (where each image pixel is given a class label) was performed using a fully-
convolutional network (FCN) adapted from the VGG-16 (available via Github as FCN-8S).
The network was trained using an Nvidia GTX-1070 8GB graphics card on 63 (48 training
and 15 validation) 200x200 pixel hand-labelled images, collected across multiple orchards
under a range of typical conditions. Commercial orchard field trials gave an average pick
cycle time of 5.5s/fruit, of which 3.0s was for the detection step. Some 89.6% of the pickable
kiwifruit could be detected. This performance was reported as significantly better than other
comparable automated systems. Wang et al. [20] provide a general review of ground-based
machine vision for weed detection that included CNN deep learning-based approaches for
detection and segmentation in the field. Many of the key challenges indicative of agricultural
applications are identified in this work, such as occlusion and overlap of leaves, varying
lighting conditions and the effects of different growth stages. Palacios et al. [21] describe a
system for quantification of grapevine flowers using a field based mobile platform for early
crop yield forecasting. Here artificial illumination and night-time operation offered some
limited structuring. A deep fully-convolutional neural network SegNet architecture (two deep
neural networks) with a VGG-19 network as the encoder were employed for semantic
segmentation. As with other examples, images were also pixel-wise manually labelled and
image augmentation (rotations and flipping) used to increase variability in the limited training
data. As discussed above, we see that in practice the use of such data augmentation
techniques are very commonly used in an effort to cope with, or avoid the need for, large
quantities of hand labelled training images. Impressive F1 (F scores are a measure of
accuracy calculated from the precision and recall) score values of 0.93 and 0.73 were
obtained for multiscale flower segmentation. A strong correlation was reported between the
estimated number of flowers and the final produce yield. Kakani [22] take a step back and
explain how agri-tech start-ups are choosing task-specific AI and vision solutions in the
context of improving yields and the goal of sustainable food supplies by 2050. Switching
from largely static crops to moving farm animals, Nasirahmadi et al. [23] demonstrate how
deep learning can be used to detect standing and lying behaviour of pigs to better than an
average precision of 0.93 and 0.95 respectively, for monitoring animal health and welfare
under varied real-world farm conditions. Use of Faster R-CNN (faster regions with
convolutional neural network features), R-FCN (region-based fully convolutional network)
and SSD (single shot multi-box detector) methods were all explored. In related work, Hansen
et al. [16] show how deep learning achieved better than 96% accuracy in biometrically
recognising pig faces on the farm and point to advantages of facial biometrics when
compared to the use of conventional RFID ear-tags. They use both a transfer learning
approach, centred on the pre-trained VGG-Face model (based on the VGG-Very-Deep-16
CNN and trained on the Labelled (human) Faces in the Wild dataset), and a CNN model,
trained from scratch with their own artificially augmented pig face data. Class Activated
Mapping using Grad-CAM was deployed to show the regions that the network used to
discriminate between pig faces (Figure 6). In an example of the benefits of transfer learning,
it is very interesting that the VGG-Face pre-trained model performed as well as it did (91%
accuracy), given that it was trained only on human and not pig faces. In addition to the
outdoor applications mentioned above, other rapidly developing areas include the use of
visual data from drones [24]. Here multispectral, thermal and visible cameras can be used for
land surveys (soil characteristics), inspection of crops for yield prediction, disease detection,
and when to irrigate (smart irrigation), or otherwise treat and harvest.

7.2 Food processing or agriculture - post-farmgate

A greater level of environmental structuring is possible in post-farm gate processing, in the
form of material handling (e.g. conveyors) and indoor controlled lighting. However, it is the
product or, in the case of quality control, the contaminates themselves that are the main
source of the variation. This makes conventional hardcoded feature extraction a significant
challenge. Applications here may include, quality control, sorting, grading, and in packing.

Rong et al. [25] present an approach for detecting foreign objects in walnuts, pointing to the
issue of irregular shapes and complex features making the manual design of suitable feature
detectors extremely problematic. Their aim was the detection of different sized natural
foreign objects (flesh leaf debris, dried leaf debris and gravel dust) and man-made foreign
objects (paper and plastic scraps, packing material and metal parts). Two different
convolutional neural network structures were used for segmenting the nuts and detecting
foreign bodies. The proposed method was able to correctly classify 95% of the foreign
objects in 277 validation images. Here, as with the production processing applications
described below, speed of operation is an important consideration. The combined
segmentation and detection processing time was less than 50ms. Similar issues were
described by Aslam et al. [26], who offer a survey of inspection methods for leather defect
detection with various deep learning architectures compared. Nasiri et al. [27] describe an
egg sorting system using a CNN to perform a classification task with an emphasis on highly
structured illumination. They classify unwashed egg images into three classes: intact, bloody,
or broken. A pre-trained CNN, VGG-16 (with weights pre-trained on the ImageNet dataset),
was modified by replacing the last fully connected layers with a classifier block. This block
included a global average pooling layer (to minimise over-fitting through decreasing the
number of parameters), two dense layers with 512 neurons and the ReLU function, batch
normalization (to maintain all inputs of the layer to the same range), dropout (as a
regularization technique to prevent overfitting), and a final dense layer in the form of the
Softmax classifier. This final layer computed the normalized probability value of the three
classes. Once again, common augmentation techniques were used to increase the size of the
manually labelled training dataset. Performance was shown to outperform traditional machine
vision-based models, achieving an average overall accuracy of 94.84% (by 5-fold cross-
validation).

7.3 Traditional manufacturing applications

While factory-based manufacturing of man-made objects offers the best opportunity for
greater levels of environmental structuring, there are still many examples of how the
introduction of deep learning can offer significant advantages.
Würschinger et al. [28] consider the application of deep learning in a range of more
traditional machine vision applications involving manufacturing production lines. In these
applications deep learning offers flexibility in terms of adaptation to differing feature
extraction tasks using transfer learning, allowing for a low-cost and fast set-up. Transferred
features can be effective for object recognition, subcategory recognition, and so lend
themselves to this kind of domain adaptation. They review a range of representative examples
where a classifier is used to detect defects, including in castings and printed circuit boards
and for cutting tool wear. Interestingly, they point to industrial developers of machine vision
solutions being initially hesitant to invest in deep learning projects, aware of the challenges
and risks but also the great opportunities. They also identify the burgeoning added potential
for deep learning systems to increase the transparency of manufacturing processes and so
enhance process understanding, contributing to root cause analysis, and helping in the
evaluation of quality control results, all aimed at improving efficiency. Ren et al. [29] present
an interesting generic deep learning-based technique for automated surface inspection – a
huge application area in manufactured products. Their classifier uses features transferred
from a pre-trained deep learning network for image classification and defect segmentation.
Results on publicly available datasets, including texture and colour industrial defects in hot-
rolled strip, x-ray images of metal pipe, in welds and in timber products, outperformed that of
hand-crafted features. AI systems are sometimes applied once the traditional defect
inspection sequence completes, where their main purpose is to either classify the defects or
reduce false positives – a concept we shall further explore later. Zhihong et al. [30] propose a
deep learning method for robotic rubbish sorting, in which bottles are identified traveling on
a conveyor within a complex cluttered background. A Fast R-CNN composed of two sub-
nets: a region proposal generation (RPN) and VGG-16 model were used for object
recognition and grasp pose estimation. In related work, Bahaghighat et al. [31] report 99%
accuracy using an optimised fine-tuning method based on the VGG-19 CNN as an end-to-end
deep learning method for inspection of bottle caps, classified in three groups (normal cap,
unfixed cap, and no cap). Against the background of deep learning’s need for large datasets,
Li et al. [17] set machine vision for manufacturing in the context of big data, where
computing efficiency has become a bottleneck to implementing real-time inspection systems
for smart industries. They propose a multibranch deep model for a manufacturing inspection
application for which a common deep model structure is modified and adapted to a fog
environment - a decentralised computing structure located between the cloud and devices that
produce data. Other examples of deep learning in manufacturing applications, that point to a
wide base of application, include for quality control in the printing industry [32] and for crack
detection in welds [33] where high accuracies of 98% and 96% respectively were reported.

7.4 Medical applications

A great deal of pioneering work in applying AI to image analysis has taken place in the
medical field. Given that many of the concepts currently being explored have transferable
application to other industries, this sector is worthy of some consideration. The large body of
published work to date indicates that deep learning is having a huge impact within the
medical sector, for medical imaging, medical data analysis, medical diagnostics and in wider
healthcare. Previously, medical applications of machine vision have largely been limited to
image processing, in which an image is enhanced in some way and then manually interpreted
by a trained clinician. However, this requires a great deal of skill, is costly, and can be
subjective and subject to error. Researchers have therefore started to explore how the image
interpretation process might benefit from AI, potentially offering improved diagnostic
performance.
Ayan et al. [34] use two well-known CNNs, Xception and VGG-16 with transfer learning and
fine-tuning in the training stage for diagnosing of pneumonia from chest x-rays. It is notable
that the Xception network was found to be more successful for detecting pneumonia cases,
while the VGG-16 network was better at detecting normal cases, indicating that different
networks have their own special capabilities on the same dataset and so the need for expertise
in selection (see section 5.2.1 above). Litjens et al. [35] provide a general survey of deep
learning in medical image analysis for organ detection, classification, segmentation,
registration, and other tasks. Importantly, they, and others [36], point to future interest in
unsupervised learning, such as generative adversarial networks (GANs). GANs use two
competing CNNs - one generating artificial data samples and the other discriminating
artificial from real samples, and can be trained to learn representative features in an
unsupervised manner. This allows existing unlabelled data to be exploited more easily by
leveraging the accessibility to big data, and so potentially has wide application well beyond
that of medicine. The cost, time and effort involved in obtaining large sets of labelled data is
one of the main limitations in deep learning – as we discussed above. Lundervold et al. [37]
provide an overview of deep learning applied to analyse medical images, with a particular
focus on one of the main areas – the MRI processing chain, from acquisition to image
retrieval, and from segmentation to disease prediction. The authors conclude that even though
there remain many challenges, the introduction of deep learning in clinical settings has
quickly produced valuable results that are reflected in a large body of high-impact
publications. One such example is that of Akkus et al. [38], who provide a review of deep
learning for MRI brain segmentation, concluding how such techniques are outperforming
previous state of the art classical machine learning algorithms. Liu et al. [39] also review
deep learning in medicine, this time for the analysis of medical ultrasound images. They
consider applications in biometrics, diagnosis, image guided intervention and therapy.
Supervised deep models are already becoming widely used for the classification,
segmentation, and detection of anatomical structures in medical ultrasound images, where
CNNs and RNNs are the two most popular architectures. Low image quality, another issue
which occurs across many applications, for example from poor target contrast, moving, non-
rigid organs and the limited availability of medical labelled training data across multiple
imaging modalities (such as MRI, X-ray, and ultrasound) are all major challenges typical of
many medical applications. Application to 3D images is also seen as an important future area
for development in this context. Inhomogeneity, varied intensity ranges, poor registration and
contrast, together with noise, are all situations where the application of conventional
techniques for pre-processing can help and we consider below other application examples of
where conventional machine vision and deep learning techniques can benefit from a close
integration.

8. Combining conventional machine vision with deep learning – the continuing

important role of conventional techniques
Next, we consider recent examples across a range of application sectors where there has been
an emphasis on the benefits of combining conventional machine vision techniques to improve
the performance of deep learning. We have already seen one major example of this in the
widespread use of image augmentation. Here we consider others, including for image pre-
processing.

A study by Xie et al. [40] had the aim of combining a convolutional neural network with
conventional computer vision techniques to automatically recognise and locate bones in
Atlantic salmon and to explore the impact of different image quality (obtained using differing
levels of image compression) on performance, in what is considered a safety critical
application. A Faster-RCNN object detection algorithm with three different convolutional
neural network models (Alexnet, VGG-16 and VGG-19) were studied. An image
compression ratio of 25% was identified as maximising costs and benefits with respect to
detection accuracy, equipment prices and detection speed. Jiang et al. [41] describe the use of
hyperspectral data for the detection of pesticide residue in post-harvest apples. They also fuse
deep learning in the form of an Alex-Net-CNN with conventional techniques by using an
Otsu segmentation and a Hough space transformation to first establish region of interest
masks, centred on the apples. Detection accuracy was better than 95% and compared
favourably with traditional k-nearest neighbour and support vector machine (SVM)
classification algorithms. Related work for online fruit sorting using deep learning to detect
internal mechanical damage of blueberries using hyperspectral data is presented by Wang et
al. [42]. Another example implementation that combines conventional machine vision
techniques with deep learning models for rapid defect identification in industrial process line
inspection was presented by Wang et al. [43]. Their factory setting allowed the use of a fixed
object pose and structured lighting to help reveal contamination defects in bottles. After
applying an image processing stage in the form of a Gaussian filter and (Canny) edge
detector, to generate an edge graph, a Hough transform was utilised to reduce the deep
learning defect detection task complexity by establishing regions of interest to which a
‘lightweight CNN’ was applied. This illustrates how the availability of big data in the form of
large numbers of example defects and defect free samples in many manufacturing
applications lends itself to a deep learning solution. Furthermore, as with many
manufacturing quality control applications of machine vision, the emphasis here was on
efficiency in terms of speed of operation in a trade off with acceptable detection accuracy.
Overall accuracy was reported to be as high as 99.6%, with a cycle time of just 47.6ms,
allowing inspection of 21 products per second. These results compared favourably both with
traditional feature based shallow machine learning, and other deep learning solutions. Li et al.
[44] present a hybrid of conventional techniques and deep learning for detecting a wide range
of defect types (scratches, floaters, light stains, and dark stains) on mobile phone screens
during production. Conventional techniques were deployed in a pre-examination stage and
regions of interest containing defect targets established using shape-based template matching.
A two-stage approach, in which several deep learning models (VGG, ResNet, GoogLeNet,
ResNeXt, SeNet, NasNet) were explored; all of which reached an impressive 99% detection
accuracy, with the simpler VGG model achieving a cycle time of 4.56s and 2.47s on a CPU
and GPU, respectively. Tang et al. [45] present an electrical component recognition method
based on deep learning and conventional machine vision techniques. Conventional pre-
processing operators, such as grayscale conversion, mean filtering, pose correction and other
techniques were used (from the OpenCV library). Component coding of different types and
materials were recognised by the CNN, with recognition results that compared favourably
with traditional techniques across a wide range of components. Jang et al. [46] also propose a
surface defect inspection solution but with an emphasis on a small training dataset in the
context of a lack of defect sample images - typical of some manufacturing applications. Their
proposed method for wafer/PCB defect inspection, a major application area from machine
vision, exploited conventional defect inspection techniques to estimate a defect probability
image. That, together with the original grey level image, formed the input to a CNN. Their
integrating of conventional inspection techniques with a CNN model was shown to be highly
effective for conventional defect inspection problems.
9. Future key challenges - promises and opportunities
The literature demonstrates that machine vision systems are being used across a plethora of
industries for a wide range of tasks. Of these, quality assurance and inspection form the
largest segment and this alone is expected to grow at a rate of 9.5% over the next four years
[47]. New markets are also developing, and given the new capabilities offer by deep learning,
there is a particular growing demand in outdoor applications. These evolving applications
mean that the global machine vision system market is expected to rapidly grow from $8.6
Billion in 2020 to $17.7 Billion by 2027 [47], accelerating the demand for improved
performance and new capabilities.

We have seen how the application of deep ANNs in industrial machine vision tasks can use
biologically informed modelling of the vision system to achieve substantial leaps in
performance. Deep learning can also help automate much of the model creation steps during
development, and in application can be more robust and flexible, more adaptable to change
and offer greater generality in application. However, it is also apparent that classical machine
vision still has an important role to play. While some tasks are not suited to deep learning
[48], for those that are, we have seen how classical techniques can often be combined with
deep learning to substantially improve performance.

However, it is also clear from the literature that there are challenges remaining, and it seems
appropriate to attempt to prioritise these as opportunities for future research. Firstly, deep
learning necessitates a large amount of, usually labelled, training data. This may not always
be accessible and there can be a scarcity of publicly available data for training CNNs. This
means, as we have seen, that in practice few researchers train deep CNNs from scratch.
Instead, many use networks that have been pre-trained on large-scale image data, where the
trained weights are applied as a feature extractor to a smaller dataset in the solution domain.
We have seen how this can be achieved by removing the last few layers of the pre-trained
CNN. This allows a system developed for one application domain to be relatively easily
transferred to another (e.g. human face to pig face biometric recognition [16]). However, this
prompts the question as to whether future research should focus on devising new data
augmentation methods for expanding often limited available data, or on how better to acquire
real big data for each task, or will transfer learning be sufficient? Secondly, datasets have to
be labelled, and in some applications (e.g. medical) by domain experts. We have also seen
how this problem can be partly solved by leveraging effective data augmentation (and to a
lesser extent automated annotation) techniques. Perhaps a more promising area of research,
aimed at addressing both limited training data and the need for manual labelling, could be to
explore methods for unsupervised learning, such as that already offered by generative
adversarial networks (GANs). As we have seen, GANs offer an unsupervised form of
learning, in which a generator network works in partnership with a discriminator network to,
for example, mimic a human expert. Unsupervised learning avoids the need for manual data
labelling by automatically discovering patterns in the data such that the network model can
generate new outputs that plausibly could have been drawn from the original real dataset. In
this regard, it is rather like generating augmented data. Impressive examples of the latter
exist, where GANs have been used to create very realistic, but completely artificial, fake
human faces [49]. Unsupervised learning and the ability to create realistic training data with
wide variation could have a huge impact across many industrial applications where training
data are limited.
It is widely recognised that deep learning is now the state-of-the-art in machine learning for
machine vision and is being increasingly widely deployed across industrial applications. We
have seen how conventional machine vision techniques can support deep learning
applications and have highlighted some of the key challenges that will need to be addressed
in the design and development of future CNN based solutions.

References
1. Y. LeCun, O. Matan, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, L.
D. Jackel and H. S. Baird: Handwritten Zip Code Recognition with Multilayer Networks, in
IAPR (Eds), Proc. of the International Conference on Pattern Recognition, II:35-40, IEEE,
Atlantic City, invited paper, 1990
2. Y. LeCun, Y. Bengio, Convolutional networks for images, speech, and time series. The
handbook of brain theory and neural networks, 3361(10), 1995
3. Y. LeCun, L. Bottou, Y. Bengio and P. Haffner: Gradient-Based Learning Applied to
Document Recognition, Proceedings of the IEEE, 86(11), 1998
4. A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet Classification with Deep
Convolutional Neural Networks, Advances in Neural Information Processing Systems 25(2),
DOI: 10.1145/3065386, 2012
5. A. Sioma, Automated Control of Surface Defects on Ceramic Tiles Using 3D Image
Analysis, Materials, 13(5), 2020
6. M. L. Smith, L. N. Smith, Dynamic Photometric Stereo - A New Technique for Moving
Surface Analysis, Image and Vision Computing, 23(9), 2005
7. B. L. Steward, J. Gai, L. Tang, The use of agricultural robots in weed management and
control, Agriculture and Engineering Biosystems Publications, Iowa State University,
DOI:10.19103/AS.2019.0056.13, 2019
8. L. N. Smith, A. Byrne, M. F. Hansen, W. Zhang, M. L. Smith, Weed classification in
grasslands using convolutional neural networks, Applications of Machine Learning, SPIE
Optics Photonics, DOI.org/10.1117/12.2530092, 2019
9. N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B. Gotway, J.
Liang, Convolutional neural networks for medical image analysis: Full training or fine
tuning?, IEEE Transactions on Medical Imaging, 35(5), 2016.
10. D. Erhan, Y. Bengio, A. Courville, P. A. Manzagol, P. Vincent, S. Bengio, Why does
unsupervised pre-training help deep learning? Journal of Machine Learning Research, 11,
Online: http://dl.acm.org/citation.cfm?id=1756006.1756025, 2010
11. K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image
recognition, arXiv 1409.1556, 09, 2014
12. K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition, 2016
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016
13. A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto,
H. Adam, Mobilenets: Efficient convolutional neural networks for mobile vision applications,
CoRR, abs/1704.04861, Online: http://arxiv.org/abs/1704.04861, 2017
14. F. Chollet, Xception: Deep learning with depthwise separable convolutions, IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
DOI.org/10.1109/CVPR.2017.195, 2017
15. A. Nasiria, A.Taheri-Garavand, Y. Zhang, Image-based deep learning automated sorting
of date fruit, Postharvest Biology and Technology, 153, 2019
16. M. F. Hansen, M. L. Smith, L. N. Smith, M. G. Salter, E. M. Baxter, M. Farish, B.
Grieve, Towards on-farm pig face recognition using convolutional neural networks,
Computers in Industry, 98, 2018
17. L. Li, K. Ota, M. Dong, Deep Learning for Smart Industry: Efficient Manufacture
Inspection System with Fog Computing, IEEE Transactions on Industrial Informatics, 14(10),
2018
18. H. Yang, S. Kumara, S. T.S. Bukkapatnam, F. Tsung, The internet of things for smart
manufacturing: A review, IISE Transactions, Vol. 51, Iss. 11, 2019
19. H. A. M. Williams, M. H. Jones, M. Nejati, M. J. Seabright, J. Bell, N. D. Penhall, J. J.
Barnett, M. D. Duke, A. J. Scarfe, H. S. Ahn, J. Y. Lim, B. Macdonald, Robotic kiwifruit
harvesting using machine vision, convolutional neural networks, and robotic arms,
Biosystems Engineering, 181, 2019
20. A. Wang, W. Zhang, X. Wei, A review on weed detection using ground-based machine
vision and image, Computers and Electronics in Agriculture, 158 2019
21. F. Palaciosa, G. Buenoc, J. Salidoc, M. P. Diago, I. Hernández, J. Tardaguilaa,
Automated grapevine flower detection and quantification method based on computer vision
and deep learning from on-the-go imaging using a mobile sensing platform under field
conditions, Computers and Electronics in Agriculture, 178, 2020
22. V. Kakani, V. H. Nguyen, B. P. Kumar H. Kim, V. R. Pasupuleti, A critical review on
computer vision and artificial intelligence in food industry, Journal of Agriculture and Food
Research, 2, 2020
23. A. Nasirahmadi, B. Sturm, S. Edwards, K-H. Jeppsson, A-C. Olsson, S. Müller, O.
Hensel, Deep Learning and Machine Vision Approaches for Posture Detection of Individual
Pigs, Sensors, 19(17), 2019.
24. D. Liuzz, G. Silano, F. Picariello, L. Iannelli, L. Glielmo, L. De Vito, P. Daponte, A
review on the use of drones for precision agriculture, IOP Conference Series Earth and
Environmental Science, 275, DOI: 10.1088/1755-1315/275/1/012022, October 2018
25. D. Rong, L. Xie, Y. Ying, Computer vision detection of foreign objects in walnuts using
deep learning, Computers and Electronics in Agriculture, 162, 2019
26. M. Aslam, T. M. Khan, S. S. Naqvi, G. Holmes, R. Naffa, On the Application of
Automated Machine Vision for Leather Defect Inspection and Grading: A Survey, IEEE
Access, DIO:10.1109/ACCESS.2019.2957427, 7, 2019
27. A. Nasiri, M. Omid, A. Taheri-Garavand, An automatic sorting system for unwashed eggs
using deep learning, Journal of Food Engineering 283 (2020)
28. H. Würschinger, M. Mühlbauer, M. Winter, M. Engelbrecht, Project Administration, N.
Hanenkamp, Implementation and potentials of a machine vision system in a series production
using deep learning and low-cost hardware, Procedia CIRP, 90, 2020
29. R. Ren, T. Hung, and K. C. Tan, A Generic Deep-Learning-Based Approach for
Automated Surface Inspection, IEEE Transactions on Cybernetics, 48(3), 2018
30. C. Zhihong, Z. Hebin, W. Yanbo, L. Binyan, L. Yu, A Vision-based Robotic Grasping
System Using Deep Learning for Garbage Sorting, Proceedings of the 36th Chinese Control
Conference, China, July 2017
31. M. Bahaghighat, F. Abedini, M. S’hoyan, A. Molnar, Vision Inspection of Bottle Caps in
Drink Factories Using Convolutional Neural Networks, IEEE 15th International Conference
on Intelligent Computer Communication and Processing (ICCP), Cluj-Napoca, Romania,
DOI:10.1109/ICCP48234.2019.8959737, 2019
32. J. Villalba-Diez, D. Schmidt, R. Gevers, J. Ordieres-Meré, M. Buchwitz, W. Wellbrock,
Deep Learning for Industrial Computer Vision Quality Control in the Printing Industry 4.0,
Sensors, 19, 3987, DOI:10.3390/s19183987, 2019
33. R. Moreno, E. Gorostegui-Colinas, P. López de Uralde, A. Muniategui, Towards
Automatic Crack Detection by Deep Learning and Active Thermography, International
Work-Conference on Artificial Neural Networks, IWANN, Advances in Computational
Intelligence, DOI: 10.1007/978-3-030-20518-8_13, 2019
34. E. Ayan and H. M. Ünver, Diagnosis of Pneumonia from Chest X-Ray Images Using
Deep Learning, 2019 Scientific Meeting on Electrical-Electronics & Biomedical Engineering
and Computer Science (EBBT), Turkey, DOI: 10.1109/EBBT.2019.8741582,2019
35. G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A.W.M.
van der Laak, B. van Ginneken, C. I. Sánchez, A survey on deep learning in medical image
analysis, Medical Image Analysis, 42, 2017
36. M. Ruppel, R. Persad, A. Bahl, S. Dogramadzi, C. Melhuish, and L. N. Smith, NANCY:
Combining Adversarial Networks with Cycle-Consistency for Robust Multi-Modal Image
Registration, International Journal of Computer and Information Engineering, 14.8, 2020

37. A. S. Lundervold, A. Lundervold, An overview of deep learning in medical imaging

focusing on MRI, Z MedPhys, 29, DOI.org/10.1016/j.zemedi.2018.11.002, 2019
38. Z. Akkus, A. Galimzianova, A. Hoogi, D. L. Rubin, B. J. Erickson, Deep Learning for
Brain MRI Segmentation: State of the Art and Future Directions. J Digit Imaging, DOI:
10.1007/s10278-017-9983-4, 2017.
39. S. Liu, Y. Wanga, X. Yang, B. Lei, L. Liu, S. Xiang Li, D. Ni, T. Wang, Deep Learning
in Medical Ultrasound Analysis: A Review, Engineering, 5, 2019
40. T. Xie, X. Li, X. Zhang, J. Hu, Y. Fang, Detection of Atlantic salmon bone residues using
machine vision technology, Food Control, DOI.org/10.1016/j.foodcont.2020.10778730, 2020
41. B. Jiang, J. He, S. Yang, H. Fu, T. Li, H. Song, D.He, Fusion of machine vision
technology and AlexNet-CNNs deep learning network for the detection of postharvest apple
pesticide residues, Artificial Intelligence in Agriculture, 1, 2019
42. Z. Wang, M. Hu, G. Zhai, Application of Deep Learning Architectures for Accurate and
Rapid Detection of Internal Mechanical Damage of Blueberry Using Hyperspectral
Transmittance Data, Sensors, 4, DOI: 10.3390/s18041126, 2018
43. J. Wang, P. Fu, R. X. Gao, Machine vision intelligence for product defect inspection
based on deep learning and Hough transform, Journal of Manufacturing Systems, 51,
DOI:10.1016/j.jmsy.2019.03.002, 2019
44. C. Li, X. Zhang, Y. Huang, C. Tang, S. Fatikow, A Novel Algorithm for Defect
Extraction and Classification of Mobile Phone Screen Based on Machine Vision, Computers
& Industrial Engineering, 146, DOI:10.1016/j.cie.2020.106530, 2020
45. H. Tang, J. Chen, X. Zhen, Component recognition method based on deep learning and
machine vision, ICIGP '19: Proceedings of the 2nd International Conference on Image and
Graphics Processing, DOI.org/10.1145/3313950.3313962, 2019
46. C. Jang, S. Yun, H. Hwang, H. Shin, S. Kim, Y. Park, A Defect Inspection Method for
Machine Vision Using Defect Probability Image with Deep Convolutional Neural Network,
Computer Vision – ACCV 2018, DOI:10.1007/978-3-030-20887-5_9, 2019
47. Machine Vision Systems - Global Market Trajectory & Analytics, Global Industry
Analysts, Inc, ID: 338506, July 2020
48. J. Bier, Is deep learning the solution to all computer vision problems?, Vision System
Design, 1 March 2019
49. Horev, Rani. "Style-based GANs–Generating and tuning realistic artificial faces."
LyrnAI: Deep Learning Explained. Available at: www.lyrn.ai/2018/12/26/a-style-based-
generator-architecture-for-generative-adversarial-networks (accessed 29 January 2021), 2018

End of Paper
888888888888888888888888888888888888888888888888888888888888888888888888888