MA-YOLO: Morphology-Adaptive YOLO for Parameter-Efficient PCB Defect Detection

Automated defect detection on Printed Circuit Boards (PCBs) is essential for electronic product quality control, but it remains challenging because PCB defects are small, morphologically diverse, and easily submerged by dense circuit textures. Fine-grained edge cues are also weakened during feature downsampling. To address these challenges, this paper proposes Morphology-Adaptive YOLO (MA-YOLO), a lightweight PCB defect-detection model based on YOLO11n. First, a Morphology-Adaptive Normalized Wasserstein Distance (MA-NWD) loss is designed to reduce the excessive sensitivity of IoU-based regression losses to pixel-level offsets, thereby improving the stability of bounding box regression for small defects. Second, the High-Frequency Integrated C3k2 (HFI-C3k2) module is designed to enhance the representation of high-frequency features associated with subtle edge defects, such as burrs and minor notches, further improving defect boundary perception under complex background conditions. Finally, Morphology-Scale-Aware Exponential Moving Average Slide Loss (MS-EMASlideLoss) is proposed by introducing a morphological scale coefficient into the original sample quality evaluation mechanism. This loss alleviates the insufficient emphasis on tiny and elongated defects during classification training and assigns higher supervisory weights to hard-to-classify defect instances. Experiments on the public PCB defect benchmark DeepPCB show that MA-YOLO achieves an mAP50 of 97.5% and an mAP50–95 of 77.9%, outperforming the baseline YOLO11n by 1.2 and 4.6 percentage points, respectively. Meanwhile, MA-YOLO retains the lightweight characteristics of YOLO11n, with only 2.565 M parameters and 6.2 GFLOPs. Compared with representative detection methods, MA-YOLO achieves a better balance between detection accuracy, model complexity, and inference efficiency. These findings suggest that MA-YOLO provides a competitive and deployable solution for lightweight PCB defect detection in industrial inspection scenarios.