于2019年11月18日发表于remote sensing上
Abstract:提出传统方法的缺点:
然后是deep learning algorithms的缺点
本文的目的:In order to enhance the feature extraction ability of the network,
方法:
-
对SAR图像进行多级解析优化,处理杂波和旁瓣,增强对SAR图像特征的分辨能力
-
propose a novel split convolution block (SCB)) to enhance the feature representation of small targets,即将SAR图像分割成小的子图像。为什么采用这个操作,Introduction中提到 Due to the overall priority in visual information processing, it is easy to lose the details of the small targets。SCB的作用: SCB divides the SAR images into smaller sub-images as input to the network, and then
-
a spatial attention block (SAB) is embedded in the feature pyramid network (FPN) to reduce the loss of spatial information, during the dimensionality reduction process
这就是说,这篇文章有个特点是,它并非单纯在使用deep learning?
在Introduction中作者有这样一句话:
To enhance the performance of CNNs, researchers have mainly investigated three important factors of the network: depth [27], width [28–30] and cardinality [31]. 这是对现有CNN网络发展的一个总结,至于是否全面,需要自己的脑子想咯
主要看2.3节
2.2. Background on RetinaNet
介绍了FPN网络和Focal loss,并介绍了RetinaNet的网络结构。
2.3. The Improvements on RetinaNet
Multi-Scale Anchor Design
文章中提出了为什么要使用anchor
- 将定位问题转换为回归问题
- 可以解决同一区域内多个物体造成的检测漏检率高的问题。
Anchorconvertstheobjectdetection problem into the regression problem between each proposal and its assigned object box [45]. By this way, the problem of high detection miss rate caused by multiple objects in the same area can be solved
为什么要重新设计anchor
The SAR ship detection dataset used in this paper is of multi-resolution, and different types of ships have different sizes, so that the ship in the chips has various sizes. In addition, the SAR images are affected by the imaging mode, and the geometric deformation caused by the incidence angle will also result in the change of the ship shape in chips. Here, we use the multi-layer feature maps for ship detection,andweneedtodesignthemultipleanchorsizesfortheSARshipdetectiondataset.
如何设计anchor:统计了边界框大小在数据集中的分布,并考虑了边界框的综合大小和纵横比
We count the distribution of the bounding box size in the dataset and take into account the comprehensive size and the aspect ratio of the bounding box.
设计anchor的原则,本文也达到了这个要求
These anchors can cover most bounding boxes in the dataset, and will not result in a huge model with too many anchors.
Split Convolution Block (SCB)
目的:提高对小目标的特征提取
方法:将原始图片分为许多大小相同的子图像来作为图像的输入。并且为了避免在切片边缘目标带来的目标特征的不连续性,仍旧使用了一个正常卷积模块,保留完整的SAR图像作为输入。于此同时,保持对输入图像整体特征的关注。
文章中说到,
!!!!!天,我第一次在文章中看到这种表述
具体做法
In this paper,we cut the input image into four sub-images of the same size, namely, the cutting positions are the middle of the distance and azimuth dimensions.
具体模块
这里的layer1_1和1_2是如何设置的???这个模块的实现方法非常简单,但是它涵盖的思想值得学习。文中也提到了两篇相关的文章,但是没有找到其获取方法,正在联系作者,希望有答复把。
Spatial Attention Block (SAB)
这个模块之前见过,就不多看了
Results and Discussions
直接看实验结果,基准模型是RetinaNet.最近看到的RetinaNet的相关文献还是比较多的,开学之前有时间可以把它们放在一起比较一下吧。不过这里给自己留一个突然想起的思考点:Retinanet实际在网络上近似于FPN,只是在损失函数方面做了改进。那么这个近似,哪里有不同。这个点值得注意。其次,这里的前馈网络中使用的都是resnet,那么是不是应该做一个实验,如果将前馈网络中替换为VGG,Inception等网络有什么结果。
这篇文章值得一读,相比于其它在网络结构上的改进,这里的思想很值得学习