论文阅读笔记：Learning Deep Features for Discriminative Localization

Introduction

Task

基于弱监督学习的图像分类和定位(检测)
相关工作:

• 弱监督目标定位
• 可视化CNN

Method

Class Activation Mapping(CAM)

CAM技术详细且简洁地展示了如何用CNN进行目标定位(检测)以及可视化，原理很简单，主要基于global average pooling(GAP)

• Firstly, get the last convolutional layer feature maps $f_k(x,y)$fk​(x,y)，is the $kth$kth channel feature map, channel num is $n$n

• Sencondly，use global average pooling to get $F_{k}$Fk​
  $F_{k} = \sum_{x,y} f_k(x,y)$Fk​=x,y∑​fk​(x,y)

• Thirdly，use a FC layer，get class score $S_{c}$Sc​，it can be used to compute softmax cross entropy loss and then to train
  $S_{c} = \sum_{k} w_{k}^{c} * F_{k}$Sc​=k∑​wkc​∗Fk​

• Finally，we can get class activation map by the weights $w_{k}^{c}$wkc​ for every class $c$c，the resolution of $M_c(x,y)$Mc​(x,y) and $f_k(x,y)$fk​(x,y) is same, and we can upsample it to get final map(size is same with oral image)
  $M_c(x,y) = \sum_{k}w_{k}^{c} * f_{k}(x,y)$Mc​(x,y)=k∑​wkc​∗fk​(x,y)

Experiments

classification result

Compared with original network(VGG、GoogleLeNet et al)，use GAP there is a small drop of 1%-2%.

Localization

Compared with fully-supervised methods，use CAM there is a large difference，at last this method not use bounding box.

Conclusion

• It is important that we can use classification-trained CNNs to learn to localization，without using any bounding box.
• The class activation mapping method is easy to transfer to other task for example captioning、VAQ et al.