Dual-perspective filter pruning via diversity and independence collaboration

While filter pruning has become a prevalent strategy for compressing convolutional neural networks, prevailing approaches predominantly rely on unilateral criteria-either intra-channel variance or inter-channel correlation-limiting their flexibility and performance retention. To bridge this gap, we propose dual-perspective filter pruning (DPFP), a framework that synergistically integrates two novel metrics: diversity-aware filter selection (DFS) and independence-aware filter selection (IFS). DFS identifies filters with low feature richness, whereas IFS targets those with structural redundancy. By fusing these complementary perspectives into a unified importance score, DPFP achieves comprehensive filter evaluation. Extensive experiments demonstrate that our approach consistently surpasses state-of-the-art methods in both compression efficiency and accuracy preservation.