Scribd
Upload
14 views

denoising_autoencoders

Uploaded by

skn.frost
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
14 views

denoising_autoencoders

Uploaded by

skn.frost
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 13

Denoising Autoencoder(DAE)

Presented by
M D Midhun Reddy
1901EE34
Introduction
● An autoencoder is an unsupervised deep learning neural
network that that learns to reconstructs the original input
through encoding-decoding process.
● A denoising autoencoder is an extension of autoencoder that is
used to remove noise from input data.
● Then model is trained to predict the original input from this
noisy input.
Architecture of DAE
● Denoising Autoencoder (DAE)
consists of two main
components:
1. Encoder Network
2. Decoder Network.
● The noise signal ~x is added to
original signal x and the
generated signal is passed to
encoder.
● X (generated signal) =
x(original signal) + ~x(noise)
Architecture of DAE
● Encoder Network : The encoder
network is responsible for
mapping the input data to a lower-
dimensional representation, also
known as a compact or bottleneck
representation.
● Decoder Network : The decoder
network takes the latent
representation as input and maps
it back to the original input space
Limitations of Autoencoder
● Autoencoders are classified based on the dimensions of
latent representation
● Undercomplete autoencoder(dim of latent < dim of
input)
● Overcomplete autoencoder (dim of latent >=dim of
input)
● Overcomplete autoencoders are prone to just copying the
input data rather than learning to capture relevant features
from input data
● Undercomplete autoencoders are not flexible because we
need to know the dimensions of latent representation
● By adding noise to the input data,autoencoders are forced
to capture the important characteristics of input data
irrespective of the dim of latent representation.
Training
● Inputs for the DAE are noisy data(images,audio etc..)
● loss function (mean square error)

● Here yi - reconstructed output,y^i = original clean signal.


● The DAE network then adjusts its weights to minimize the loss
function.
Applications of DAE
Denoising autoencoder has a broad range of applications in
various fields, where the removal of noise and the clear data is
crucial.
Object detection:
● Object detection is the task of identifying and locating
objects of interest in a video.
● By removing noise, DAEs can improve the accuracy of
object detection algorithms and make it easier for human
operators to monitor and interpret the video.
Applications
Data compression:

● Denoising autoencoders can also be used for data


compression, where the encoder encode the input data into a
lower-dimensional latent space.
● This can be useful for reducing the storage requirements of
large datasets, without sacrificing important information

Speech Denoising :

Denoising autoencoders are also used for removing noise


from speech signals, which can improve the performance of
speech recognition systems.
Applications of Denoising Autoencoder in
Video Surveillance
Image Denoising

DAEs can be used to remove noise


from video frames.
● DAEs are trained for noisy
images
● DAEs output is compared with
original clean images
● DAEs weights are adjusted to
minimize the loss
function(mean square error).
Anomaly Detection :

● Denoising autoencoders can be used


to detect anomalous behavior in video
surveillance footage.
● Main features for the model are
skeleton pose and motion.
● The DAE are trained on normal
behavior patterns(walking).
● This anomaly is detected by using his
unusual skeleton pose and motion
compared to those of normal
activities.
● Skeletons in red denote high anomaly
scores(anomaly score = reconstructed
error)
Challenges
1. High computational requirements
Denoising autoencoders require significant computational resources,
especially for processing large amounts of video data in real-time. This can
be a challenge for video surveillance systems that need to process video
feeds in real-time.
2. Overfitting
Denoising autoencoders can be prone to overfitting, where the model
becomes too specific to the training data and fails to generalize to new
data. This can be a problem in video surveillance, where there is often a
high degree of variability in the input data.

3. Need for large amounts of training data:


Denoising autoencoders require large amounts of training data to learn
effective denoising strategies. In video surveillance, it can be difficult to obtain
large amounts of high-quality training data, especially for rare or unusual events.
Implementation
Denoising autoencoders for image denoising.
● Dataset for the above model is
Berkeley Segmentation Data Set and Benchmarks 500
● Dataset contains 500 images (300 train,200 test).
● Can implement image augmentation (eg: random cropping)
to increase the image samples.
● Reference : Xiao-Jiao Mao, Chunhua Shen, Yu-Bin
Yang,Image Restoration Using Convolutional Auto-encoders
with Symmetric Skip Connections
References
● https://towardsdatascience.com/denoising-autoencoders-
explained-dbb82467fc2
● https://www.youtube.com/watch?v=0n2x_D-ZmmU
(Denoising autoencoders tutorial by Prof.Mithesh M.
Khapra)
● https://ieeexplore.ieee.org/stamp/stamp.jsp?
arnumber=8953884(Learning Regularity in Skeleton
Trajectories for Anomaly Detection in Videos)
● Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., & Manzagol,
P. A. (2010). Stacked denoising autoencoders: Learning
useful representations in a deep network with a local

You might also like