Analysis of Filter Index Alignment in Image Super-Resolution Pruning

초록

Recent advancements in efficient image super-resolution (SR) have leveraged various compression techniques such as knowledge distillation and quantization. However, these methods often require considerable computational resources and memory. Network pruning has emerged as an effective approach to significantly reduce the number of parameters and the overall model size. In the SR domain, various studies have also explored filter pruning. Existing methods have commonly adopted filter alignment strategies that align filter indices across residual connections. However, there has been little analysis on how such alignment strategies impact the performance of SR filter pruning. In this paper, we investigate the performance effects of the filter alignment methods in SR models and provide experimental evidence to guide the choice of optimal alignment strategies. © 2025 IEEE.

키워드