CAM-based weakly supervised learning object localization device and method
Disclosed are a CAM-based weakly supervised object localization device and method. The device includes: a feature map extractor configured to extract a feature map of a last convolutional layer in a convolutional neural network (CNN) in a process of applying an image to the CNN; a weight vector binarization unit configured to first binarize a weight vector of a linear layer in a process of sequentially applying the feature map to a pooling layer that generates a feature vector and the linear layer that generates a class label; a feature map binarization unit configured to second binarize the feature map based on the first binarized weight vector; and a class activation map generation unit configured to generate a class activation map for object localization based on the second binarized feature map.
1. A CAM-based weakly supervised object localization (WOSL) device, comprising:
a feature map extractor configured to extract a feature map of a last convolutional layer in a convolutional neural network (CNN) in a process of applying an image to the CNN;
a weight vector binarization unit configured to first binarize a weight vector of a linear layer in a process of sequentially applying the feature map to a pooling layer that generates a feature vector and the linear layer that generates a class label;
a feature map binarization unit configured to second binarize the feature map based on the first binarized weight vector; and
a class activation map generation unit configured to generate a class activation map for object localization based on the second binarized feature map.
2. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the pooling layer is implemented to generate the feature vector by performing global average pooling on the feature map.
3. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the weight vector binarization unit performs the first binarization on the weight vector based on a threshold.
4. The CAM-based weakly supervised learning object localization device of claim 3 , wherein the weight vector binarization unit performs a grid search based on a plurality of relative thresholds in order to optimize the threshold.
5. The CAM-based weakly supervised learning object localization device of claim 4 , wherein the weight vector binarization unit performs the grid search using Equation 1 below:
τ w c =min( w c )+{max( w c )−min( w c )}*θ w , [Equation 1]
wherein, τ wc denotes an optimal threshold for a weight of a c-th target class, c∈C (where C is the number of target classes), w∈W (where W is the weight vector), and θ w denotes the relative threshold.
6. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the feature map binarization unit applies the first binarized weight vector to the feature map to select at least one of the channels of the feature map.
7. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the feature map binarization unit performs the second binarization on the at least one of some channels based on a feature binarize threshold.
8. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the feature map binarization unit performs the second binarization using Equation 2 below:
F k =1( F k ( x,y )> T f ), [Equation 2]
wherein, F k (x,y) denotes an element of a k-th channel of a feature map F k for a y-th row and an x-th column, and T f denotes a feature binarize threshold.
9. The CAM-based weakly supervised learning object localization device of claim 1 , wherein the class activation map generation unit generates the class activation map by aggregating the at least one of second binarized some channels.
10. A CAM-based weakly supervised object localization (WOSL) method, comprising:
extracting a feature map of a last convolutional layer in a convolutional neural network (CNN) in a process of applying an image to the CNN;
first binarizing a weight vector of a linear layer in a process of sequentially applying the feature map to a pooling layer that generates a feature vector and the linear layer that generates a class label;
second binarizing the feature map based on the first binarized weight vector; and
generating a class activation map for object localization based on the second binarized feature map.
11. The CAM-based weakly supervised object localization method of claim 10 , wherein the first binarizing includes performing the first binarization on the weight vector based on a threshold.
12. The CAM-based weakly supervised object localization method of claim 10 , wherein the second binarizing includes selecting at least one of some channels among channels of the feature map by applying the first binarized weight vector to the feature map.
13. The CAM-based weakly supervised object localization method of claim 10 , wherein the second binarizing includes performing the second binarization on the at least one of some channels based on a feature binarize threshold.