Aiming at the problems of insufficient perception of edge feature details and inadequate interaction between channel and spatial information in residual image classification networks, this paper proposes a Mach Band Mechanism-based Image Classification Network (MBMNet). The network is built on the ResNet-34 framework and designs a Dynamic Mach Band Module (DMM) and a Mach Band Attention Module (MBM). Among them, DMM performs edge detection through depthwise convolution and the Laplacian operator, and realizes dynamic balance between edge enhancement and suppression through a learnable scaling factor, thereby strengthening the module’s perception of key edges. MBM constructs a Channel Enhancement Branch (CEB) and a Spatial Enhancement Branch (SEB) in the channel and spatial dimensions, respectively, and embeds DMM into both branches. Specifically, CEB extracts key channel information through mixed pooling and depthwise separable convolution, and further strengthens edge and structural variations inside the channel together with DMM; SEB extracts local spatial textures and structural features through multiple depthwise separable convolutions, and enhances spatial edge responses in key regions under the influence of DMM. Finally, the features output by the two branches are fused through the Feature Fusion Module (FFM), enabling collaborative modeling and information interaction between channel and spatial features, and residual fusion is further applied to maintain semantic consistency and stable feature propagation. On the CIFAR-10, CIFAR-100, SVHN, Imagenette and Imagewoof datasets, MBMNet achieves classification accuracies of 96.87%, 81.08%, 97.64%, 92.25% and 85.31%, respectively, and compared with other mainstream networks, MBMNet achieves average improvements of 1.70%, 3.15%, 2.19%, 4.39% and 5.76% on these five datasets. Experimental results show that MBMNet can effectively enhance the perception of edge and structural details, promote the collaborative interaction of channel and spatial features, and thus significantly improve the accuracy and robustness of image classification.