An Improved YOLOv11-Based Model for Piston Corner Defect Detection

In industrial manufacturing, tiny defects at piston corners can cause serious problems if they are not detected accurately and in time. Such defects may affect equipment performance and even threaten operational safety. However, defect detection remains challenging due to very small target sizes, complex imaging conditions, and strong reflections on metal surfaces. To address these issues, this paper proposes a lightweight model named MDE-YOLO. First, a non-dilated multi-scale convolution module (PKI) is introduced to enhance local details and multi-scale feature representation. Second, a dynamic feature fusion module (MFM) is used to adaptively adjust the weights of features from different scales. In addition, the detection head integrates MBConv structures and SE channel attention, which improves feature extraction while keeping computational cost low. Experiments on a self-built piston defect dataset show that MDE-YOLO achieves 96.6% mAP@0.5, with a computational cost of about 5.2 GFLOPs and an inference speed of 182.7 FPS. The model has only 2.3M parameters, yet it outperforms many mainstream detection models. Further evaluation on the MS COCO dataset demonstrates that MDE-YOLO maintains good generalization ability in complex scenes. These results indicate that the proposed method is well suited for industrial defect detection tasks that require both high accuracy and real-time performance.