Breeds Classification with Deep Convolutional Neural Network
Yicheng Zhang, Jipeng Gao, Haolin Zhou
Abstract
In this work, we utilized a famous convolutional neural network structure with small convolutional filters and deep layers to distinguish different breeds of cats, and this network reached high accuracy. What is more important, this work explored what evidence neural networks depended on to identify only slightly different objects. To make our network more comprehensible, we did the visualization, including the images that each filter most wanted to see, the output images of convolutional layers, and the heat maps. By analyzing these results, we generalized the special case to ordinary cases, and explained the method convolutional neural networks use to identify features. Finally, we discussed the similarities of between how humans and convolutional neural networks see the world.