Auto-DenseNet: DenseNet Optimization Method Based on Evolutionary Algorithm

To address the issue of time-consuming and laborious design of network structure parameters and connection methods, in this paper we propose an evolutionary algorithm-based DenseNet network optimization method (Auto Dense Convolutional Network, Auto-DenseNet). Firstly, the internal parameters of DenseNet are encoded to generate an initial population, and crossover and variation operators are designed to ensure the effectiveness of the evolution of offspring. Secondly, the dataset is imported into the decoded individual evaluation fitness. And then, the superior individuals are selected to enter the next iteration. Finally, the optimized network individuals are selected for the image classification task. After tuning validation on the MNIST-RD dataset, tests were conducted on three widely used benchmark image classification datasets. The experimental results demonstrate that the network structure searched by the Auto-DenseNet algorithm outperforms most existing structures in terms of classification performance, and the number of network parameters.