An Attention-Guided Global and Local Fusion Framework for Lesion-Focused Image Classification

To address the problems of sparse discriminative signals and background noise interference in lesion-focused image classification, researchers proposed an attention-guided global and local fusion framework. Based on DenseNet-121, this framework consists of three branches: the global branch uses gradient-weighted class activation maps (Grad-CAM) to generate attention maps and produce masked inputs; the local branch introduces the Convolutional Block Attention Module (CBAM) to extract detailed features; the adaptive fusion branch dynamically integrates global and local information through instance-specific weights. Experiments show that this framework performs better than single-branch approaches on synthetic speckle pattern datasets as well as three benchmark datasets: skin lesions, guava leaves, and grape leaves. The accuracy rate reached 97.75% on the skin lesions dataset and 99.64% on the guava leaves dataset.