Abstract:To address the issues in the field of image classification, including limited capability in extracting complex texture features, insufficient mining of local information correlations, and insufficient focus on key features, this study proposes an image classification network based on multi-scale texture perception and class-activated block attention, termed as MTCANet. MTCANet is built on the ResNet-34 residual network. First, a multi-scale texture perception convolution (MTPC) unit is designed, where convolution kernels are initialized using parameterized Gabor functions and combined with three-scale depthwise separable convolutions to enable channel-independent processing and diversified receptive fields, thus enhancing the representation of multi-scale texture features. Secondly, a block-wise class-activated feature extraction (BCAF) module is proposed. Based on the class activation mechanism, block-level and intra-block pixel-level confidence scores are computed via blocking operations, forming a dual-layer weighting mechanism that enhances salient region features while suppressing background interference, thus generating more discriminative local feature representations. Finally, a class-activated block feature interaction attention (CAPI) module is designed to model block-wise semantic interactions to enhance local detail consistency, aggregate local features into global semantics, and employ an attention mechanism for deep fusion of local and global features, thus improving target perception and discrimination in complex backgrounds. Experimental results demonstrate that MTCANet achieves classification accuracies of 89.99%, 80.12%, and 91.69% on the Imagenette, Imagewoof, and Flowers-102 datasets, respectively, with an average improvement of 2.37%, 1.90%, and 1.11% compared to existing state-of-the-art methods, confirming significant enhancement in image classification performance.