Compression Algorithm of Convolutional Neural Network Model

The exponential growth of data requires powerful analytical methods and processing tools, and deep learning demonstrates powerful analytical and predictive capabilities. CNN is excellent model in the field of deep learning, enabling powerful feature abstraction and image understanding that goes beyond the accuracy of recognition using the human eye. However, while the CNN model is constantly approaching the precision limit, the network depth and model parameters have multiplied, and the huge memory consumption and excessive storage occupation cannot be applied on many embedded and mobile devices. This paper constructs a compression algorithm model, including convolution layer, filter and pruning algorithm, and focuses on structured pruning and filter importance calculation. Using MNIST data set, the LeNet-5 model was trained by parameter pruning according to different pruning ratios, and the model compression was realized. The application of the research results in this paper can significantly reduce the redundant parameters in the network model, thus achieving the purpose of reducing the computing cost and storage space, and promoting the development of deep learning.