Detailed explanation of deeplab series (simple and practical annual summary)

Hello everyone , I meet you again , I'm your friend, Quan Jun .

1、 Why convolution neural network has translation invariance ？

Invariance can be divided into ：

• Translation invariance ：Translation Invariance
• rotate / Perspective invariance ：Ratation/Viewpoint Invariance
• Scale invariance ：Size Invariance
• Illumination invariance ：Illumination Invariance

CNN = Convolution + Pooling

The features of the image are translated , So in the process of convolution , Feature extraction is translated accordingly .

During pooling , It returns the maximum or average value in the receptive field , The receptive field still corresponds to the characteristic information of the image .

therefore ,CNN It has translation invariance .

2、 Why is it difficult for convolutional neural network to deal with pixel level classification ？

• Repeated pooling and down sampling lead to a sharp decline in resolution , Location information is lost and difficult to recover
• Spatial invariance leads to the loss of detail information

In fact, convolution and pooling lead to the loss of a lot of image detail location information .

3、deeplab v1 What has been done ？

problem ：DCNNs The invariance of is not enough for semantic segmentation .

Method ： combination DCNNs And probability graph models , namely DCNNs The last layer of response and conditional random fields solves the segmentation problem .

4、deeplab v2 What has been done ？

problem ： Semantic segmentation problem .

Method ： Cavity convolution 、 Pyramids pool 、DCNNs+CRF.

1、 Use sampling filter or null Hole convolution （Atrous Convolution） Highlight convolution , It is a powerful tool in intensive prediction tasks . Void convolution is allowed in DCNNs The calculated characteristic response shows the control resolution . At the same time, the filtering receptive field is effectively expanded to mix richer context information without increasing the amount of parameters and computational complexity .

2、 Put forward a Pyramid with empty space pooling（ASPP） Segment the target stably on multiple scales .ASPP A filter with multiple sampling rates and effective field of view is used to detect the incoming convolution feature layer , So as to capture the object and image context of multiple scales .

3、 Propose the location of the target boundary , By combining DCNNs And probability models . Ordinary DCNNs in max-pooling And mining to maintain invariance, but has an impact on positioning accuracy , use DCNNs and CRF Combine to solve the problem of positioning accuracy .

Void convolution ： Enhance intensive forecasting 、 Expand the feeling field .

Pyramidal pooling of empty space ： Multiscale image representation .( Multi scale feature extraction + Information fusion )

DCNNs+CRF effect ： Structural prediction of precise boundaries .

5、deeplab v3 What has been done ？

problem ： The accuracy of semantic segmentation .

Method ： Ed - Decoding structure , Introduce decoding module ; combination Xception model And depth separable convolution （depthwise separable convolution） and ASPP、 Decoding module .

DeepLabv3, By increasing the Simple and effective decoding module fine segmentation results Especially the object boundary . further , Use Xception model And depth separable convolution （depthwise separable convolution）, combination ASPP And the decoding module gets a faster 、 Stronger editing - Decoding network .

6、deeplab Series of work ？

Deeplabv1 Use CRF post-processing , Improve the accuracy of segmentation boundary ;

Deeplabv2 Use void convolution to expand receptive field , Using hole space pyramid pooling to realize multi-scale prediction and context information extraction , At the same time, post-processing is used CRF;

Deeplabv3 Do not use post-treatment , Use of editing - Decoding structure improves segmentation boundary prediction , Use depth to separate convolution and Xception modular .

Publisher ： Full stack programmer stack length , Reprint please indicate the source ：https://javaforall.cn/128203.html Link to the original text ：https://javaforall.cn