Summary of Convolutional Neural Network Compression Technology

Deep convolutional neural networks (DCNNs) obtain dramatically accuracy improvement in the area of computer vision for recent years. However, because of their large demand for memory, power, and computational ability, it is difficult for DCNNs to be employed in the light-weight devices such as mobile. Therefore, a natural idea is compressing the DCNNs, while reduce the demand for the memory and computational ability and not reduce the classification accuracy too much. In recent years, model compression gain a lot of improvement. And we summaries the compression methods in this paper. DCNNs mainly classify three types, feed-forward deep networks (FFDN), feed-back deep network (FBDN) and bi-directional deep networks (BDDN). The methods of compression and acceleration are roughly categorized into four schemes: parameter pruning and sharing, low-rank factorization, transfer /compact convolutional filters and knowledge distillation. In this paper, we introduce the performance, related applications, advantages and drawbacks of each CNNs and techniques for compacting and accelerating DCNNs model. In the last, we summarize the paper and propose the possible future work in model compression.