Analysis of the Influence Degree of Network Pruning on Fine-grained Image Processing Tasks
Jianrong Xu, Boyu Diao, Bifeng Cui, Chao Li, Yongjun Xu, Hailong Hong · 2021 IEEE 6th International Conference on Signal and Image Processing (ICSIP) · 2021
The network pruning algorithm has achieved significant effects in the compression and acceleration of neural network models, and has been used to reduce the inference cost of neural network models. However, the impact of the compressed model on fine-grained image analysis tasks is rarely considered in the evaluation of this algorithm. In response to this problem, this article carried out an analysis of the influence of the model after network pruning on fine-grained image processing tasks. For the selection of fine-grained image processing tasks, this article mainly takes the style transfer task based on the fine-grained features of the Gram matrix as the background. Considers the performance of the model before and after the network pruning on the task, the paper carries out relevant analysis on effect of generating image from the style transfer, different color channels, cosine similarity and etc. Experimental results show that the model after network pruning has a more streamlined model structure, which will not affect the overall effect of the fine-grained image processing, but there may be a certain loss in some details of the fine-grained image processing task. The research results of this paper show that the model after network pruning can still be well qualified for fine-grained image processing tasks, which has important reference value for the research and application of network pruning methods in fine-grained image processing tasks.