Segmented Convolutional Neural Network Pruning Method Based on Density Peak Clustering

In recent years, deep learning technology has developed rapidly, and the capabilities of models have become increasingly powerful. However, these models also have problems such as large size and computational complexity. In order to reduce model size and improve their inference speed without affecting performance, model compression is imperative, and pruning is the most widely used method for that. This article proposes a segmented pruning method based on density peak clustering (SDPC). This method first evaluates the impact of convolutional kernels on the output loss function in the network, in order to better preserve the convolutional kernels that have a significant impact on the network output results. Enhance the image recognition level of the remaining convolutional kernels after pruning, rather than relying solely on the analysis and judgment of the relatively isolated attributes inside the convolutional kernels, and eliminate the truly low contribution convolutional kernels. By segmenting the network, each segment can be analyzed to better measure the importance of convolutional kernels in shallower networks. This solves the problem of lower contribution scores in shallow networks due to the influence of overall network depth, thus better evaluating the value of shallow convolution kernels and achieving the goal of correctly pruning them. In order to further optimize the accuracy of pruning in the above process, this paper also applies density peak clustering algorithm to classify convolution kernels with different characteristics in the network based on segmented pruning, and romoves similar convolution kernels belonging to the same type. It made it possible to retain the convolution kernels that contribute the most to the accuracy of the network, while removing functionally redundant convolution kernels, achieving a better balance between network accuracy and volume.