Abstract
Image classification, as the core task of computer vision, has broad application value in fields such as medical diagnosis and intelligent transportation.However, the ability of differentiable neural architecture to search (NAS) for local information is weak, which limits the accuracy and long-distance information capture capability of the algorithm. Therefore, based on this, the study introduces visual attention mechanism and proposes an improved model that replaces the original convolution operator and adds residual structure in the macro structure to enhance the model's information acquisition ability and classification accuracy. The research results show that after 600 rounds of training on the CIFAR-10 dataset, the final accuracy of the improved model reached 97.2%. The runtime memory usage on the CIFAR-100 dataset is only 44.52%, a decrease of 44.56% compared to the baseline model. In the testing on the ImageNet dataset, the classification accuracy of the research model is 94.01, the search parameter required is only 4.8MB, the search time is shortened to 0.5d, and the minimum number of floating-point operations is 3.7G, significantly better than other mainstream algorithms. The above results indicate that the research method can effectively solve the shortcomings of traditional differentiable neural architecture search in local and remote information acquisition capabilities, providing important technical support for improving the accuracy and efficiency of image classification technology.