CASML: Combining Cross-Scale Attention and Separate Mix-Layer for Lightweight Classification Network

We propose a Separate Cross-Layer Module to achieve a lightweight method by splitting the network and a Cross- Scale Attention Module to enhance feature representation by extracting cross-scale features. By combining these two modules, we introduce CASML backbone for classification tasks. Through experiments, we demonstrate that we can reduce the computational costs by 25% and improve classification accuracy by 1.8% on the CIFAR100 dataset. These modules can be incorporated into other classification networks to achieve similar performance, demonstrating high versatility.