Abstract:To address the limitations of traditional convolutions in image classification tasks—namely their lack of explicit modeling of feature map statistical distributions, which results in insufficient saliency representation and susceptibility to background noise—a saliency-aware image classification network (SANet) is proposed. Based on ResNet-34, SANet integrates three key modules to enhance feature saliency and robustness. The saliency-aware convolution (SAConv) employs Z-score normalization and a nonlinear mapping function to strengthen salient feature representation and extract more discriminative information. The saliency-aware pooling (SP) module combines SAConv with average pooling to filter critical information while suppressing redundancy, mitigating feature loss during propagation. By fusing SAConv and SP, the saliency-aware block (SABlock) achieves inter-layer enhancement and stable transmission of salient features. Finally, the saliency-aware excitation (SAE) module incorporates global statistical information to model inter-channel contextual dependencies, enabling adaptive channel weighting and improving generalization performance. Experiments on CIFAR-10, CIFAR-100, SVHN, Imagenette, and Imagewoof achieved classification accuracies of 96.87%, 81.47%, 97.64%, 92.43%, and 85.54%, respectively, yielding average improvements of 1.88%, 3.56%, 2.20%, 3.87%, and 5.45% over mainstream networks. These results demonstrate that the proposed saliency-aware mechanism effectively enhances the network’s responsiveness to critical regions, suppresses background interference, and significantly improves both generalization and robustness.