CNN-Based Hidden-Layer Topological Structure Design and Optimization Methods for Image Classification

Convolutional neural networks (CNN) is one of the most important branches of deep learning, which always shows the excellent performance on image classification via unique convolution operations. However, the generalization ability of CNN is always limited due to lack of the specific guidelines in hidden-layer design, Kernel design and Weight initialization design. In this paper, a new topological design method is proposed by analyzing abstract edge information (called texture) in feature map based on the experimental and numerical analysis. Especially, the prior number of convolution kernels in the first layer and combinatorial optimization of all hidden layers are applied to initialize the entire network topology. The experiments based on the MNIST, Chest X-ray and CTs dataset indicate that (1) Traditional CNN layers with doubling nodes are not essential to optimize the hidden-layer topology because of the texture features that extracted from different datasets. (2) Improved hidden-layer topology of the CNN can outperform the better performance in classification-tasks and improvement up to 30% compared with the benchmark methods.