A New Channel Pruning Method Combining Lowrank Feature Optimization with Noise Adaptive

This paper proposes a novel convolutional neural network (CNN) compression method to address the computational and memory demands that hinder CNN deployment on resourceconstrained hardware. First, we introduce a pruning strategy based on channel correlation coefficients, which assigns classspecific masks to each channel. This enables efficient model compression by selectively retaining informative channels while eliminating redundant ones. Second, we design a low-rank inducing loss term derived from feature map analysis. This term encourages the network to learn with enhanced low-rank properties, thereby strengthening the linear correlation among feature maps and consequently facilitating greater redundancy removal during pruning. Compared to conventional approaches, this loss function improves performance without requiring intricate structural modifications. Furthermore, a noisy training module is incorporated during the fine-tuning stage to enhance model robustness and stability. Finally, extensive evaluations on the CIFAR-10 and CIFAR-100 datasets demonstrate the effectiveness of our approach. Evaluations utilizing VGG-16, ResNet-56, and ResNet-110 as backbone architectures show that our method consistently achieves superior compression ratios and lower computational complexity compared to state-of-the-art model compression techniques. Our codes are at https://github.com/qy513449931/RLP.