CNN Hyperparameter Optimization Based on CNN Visualization and Perception Hash Algorithm

Abstract

In this paper, the network structure and the optimal hyperparameter selection which affect the performance of the model are obtained through the analysis of the convolutional neural network model with mathematical interpretation and visualization. In the study, we used visual methods such as deconvolution and Guided Grad-CAM to display the network structure, parameter changes, and the learning process of the model convolutional layer. Simultaneously, we developed CNN hyperparameters optimization strategy based on the perceptual hash algorithm according to its training characteristics. This method significantly improves the accuracy of image classification of the model and the generalization ability of the model and also provides certain theoretical support for the optimization and understanding of deep learning models in practical application. In addition, the hyperparameter optimization method based on the deep learning model feature map reconstruction visualization proposed in this paper also provides a good idea for the formulation of model training strategies.