SDSHNet: Dynamic feature fusion with transformer and star operation for efficient detection in aluminum alloys microscopic inclusion

Accurate detection of microscopic inclusions is critical for aluminum alloy quality inspection in aerospace and precision manufacturing. However, existing deep learning based detectors often struggle to balance detection accuracy and computational efficiency when dealing with tiny, low-contrast inclusions under complex microstructural backgrounds. In this paper, we propose SDSHNet, a lightweight detection framework that emphasizes task driven dynamic feature interaction rather than architectural complexity. The proposed framework adopts Star Operation as an efficient backbone to preserve fine-grained structural details with low computational overhead. On this basis, A dynamic feature fusion module is designed to adaptively enhance feature representation in key regions, significantly improving the detection capability for tiny, low-contrast inclusions. In addition, a hierarchical attention mechanism based on HiLo attention is incorporated to enhance multi-scale representation, and Shape-IoU loss is employed to improve bounding box regression stability for small and irregularly shaped inclusions. Extensive experiments conducted on a microscopic aluminum alloy inclusion dataset demonstrate that SDSHNet achieves 90.7% mAP@0.5 and 60.9% mAP@[0.5:0.95], while maintaining substantially reduced computational complexity compared with mainstream convolutional and Transformer based detectors. Ablation studies further verify the individual and synergistic contributions of each component. These results indicate that SDSHNet provides an effective and deployment friendly solution for high throughput industrial inspection of aluminum alloy materials. The code is available at https://github.com/wtl985/SDSHNet .