Leveraging Attentional Mechanisms and Multiscale Feature Fusion for Enhanced Fine-Grained Image Classification

Fine-grained image classification is often hindered by challenges such as background interference, inaccurate key region localization, and model complexity. To address these, we propose a novel network based on YOLOv7, where we redesign the backbone with the lightweight Ghost BottleNeck module and replace conventional convolutions in the neck with GhostConv, reducing model complexity. We also introduce the parameter-free SimAM attention mechanism, which captures spatial and channel correlations to enhance relevant features and suppress irrelevant ones, improving target region localization. Additionally, a feature-selective pyramid pooling module is integrated to capture multiscale features, boosting the model’s perceptual ability. Experimental results show that our approach significantly improves classification accuracy and robustness.