文献类型：期刊文献

英文题名：Intra-group neighborhood relationship-aware channel pruning

作者：Pan, Yu[1];Chen, Ning[1];Zhu, Hongqing[1];Zhu, Zhiying[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：168

外文期刊名：DIGITAL SIGNAL PROCESSING

收录：;EI(收录号：20253919242293);WOS:【SCI-EXPANDED(收录号:WOS:001584891300002)】；

基金：This work is supported by the National Natural Science Foundation of China [Grant Numbers 61771196, 61872143] . The authors would like to thank the authors of [10] for providing codes of their method.

语种：英文

外文关键词：Intra-group neighborhood relationship; Structural pruning; Clustering-based pruning; K-Medoids clustering; Principal Component Analysis (PCA)

摘要：The high computational and memory requirements of Convolutional Neural Networks (CNNs) limit the deployment of CNN-based image processing models on edge computing devices. To this end, the clustering-based structural pruning methods have been studied to prune the redundant channels. However, the conventional clustering-based pruning methods may not achieve satisfactory performances for the following reasons. First, the clustering is performed only based on the feature of the current layer, which is not enough to identify the importance of each channel precisely. Second, the K-Means or K-Means++ clustering-based pruning methods may be affected by abnormal channels easily. Third, the strategy of pruning the channels around centroids may prune important channels. Fourth, the number of clusters needs to be set manually, which may affect the flexibility and generalization. To solve these issues, an intro-group neighborhood relationship-aware (IGNRA) channel pruning method is proposed. First, DepGraph is adopted to construct the dependency graph, based on which the global-level importance of each channel is assessed. Second, K-Medoids is adopted to perform clustering to reduce the influence of abnormal channels. Third, the centroids of clusters with multiple channels are viewed as redundant channels and pruned directly, while the centroids of those with a single channel are retained due to their unique roles in the downstream tasks. Fourth, the number of the clusters is set according to the pruning ratio automatically to enhance the method's flexibility and generalization. Extensive experimental results on 9 models for two image processing tasks on 6 datasets demonstrate that the proposed method outperforms the state-of-the-art pruning methods, and each key module contributes to the performance enhancement.