Adaptive convolutional network pruning through pixel-level cross-correlation and channel independence for enhanced model compression

In deep learning model optimization, the majority of pruning techniques employ predefined criteria to guide filter elimination. Although effective in identifying redundant filters, these methods often overlook intrinsic linear redundancy embedded within the network during early training stages, thus not fully exploiting the latent capacity of the model. To address this limitation, we propose an innovative structured pruning approach grounded in a Pixel-Level Cross-Correlation(PCC) loss term. By incorporating this data-driven loss function in the pre-training phase, our method promotes the formation of more robust intrinsic connections among feature maps. Leveraging PCC of output features, the network discerns finer inter-pixel relationships. This approach enhances feature map granularity, refining local feature representation and overall map quality. During the pruning phase, we employ a channel screening strategy based on Channel Independence (CI), which allows the classification of channels by importance and preserves adequate linear redundancy. Following pruning, the model is retrained and fine-tuned. Compared to state-of-the-art approaches, our method achieves substantial compression while retaining performance. For instance, on the Canadian Institute For Advanced Research 10 (CIFAR-10) dataset, the Visual Geometry Group 16-layer network (VGG-16) model demonstrates a 0.87% improvement in pre-training accuracy (94.59%) and achieves 93.40% accuracy post-pruning, with pruning rates of 92.75% for parameters(Params) and 84.61% for Floating Point Operations (FLOPs). Similarly, on the ImageNet(LSVRC-2012) dataset, our method reduces Params and FLOPs of Residual Network-50(ResNet-50) by 60.24% and 58.40%, respectively. This demonstrates a significant reduction in model size and computational complexity while maintaining high performance.