7 – 05 Feature Maps V1RENDER V3

Let’s start near the beginning of this CNN. We’ve talked about how the first convolutional layer in a CNN consists of a number of filtered images that are produced as the input image is convolved with a set of image filters. These filters are just grids of weights. For a convolutional layer with four filters, four filtered outputs will be produced. Each of these filtered output images is also called a feature map or activation map because each filtered image should extract certain features from an original image and ignore other information. For a given image, each of these maps will activate in some way, displaying activated bright pixels or not in each map. A good example of a filter is a high-pass filter, which when applied to an input image creates an activation map that activates the most when it sees high frequency features. You’ve programmed your own high-pass filters. One interesting thing about most CNN’s is that as the network trains, and the weights that make up each convolutional kernel are updated, the first convolutional layer often learns to create high-pass filters just like these. Next, let’s learn more about how to visualize the filters and feature maps in this first convolutional layer.

%d 블로거가 이것을 좋아합니다: