Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition

Figure 6

Each image indicates a certain convolutional layer’s output feature map. Each column consists of 5 feature maps, which are extracted during training process of 3 different VGG-11 models. From top to bottom, the feature maps are extracted after training 1, 30, 60, 90, and 120 epochs, respectively. From left to right, the output feature maps are randomly picked from different convolutional layers, which locate right before the first pooling layer in each model (left column: the first convolutional layers in model A; middle column: the third convolutional layer in model B; and right column: the fifth convolutional layer in model C). Therefore, the kernel size of all maps is 32×32. The colour bar on the right of each map indicates the pixel intensity; i.e., the brighter the pixel colour, the bigger the pixel value. In addition, black corresponds to value 0.