YOLO-BiNWD: Small Object Detection Algorithm for Screen Defect Based on Improved YOLOv10

Abstract

Surface defect detection on smartphone screens is critical for ensuring product quality and user experience, but traditional methods often struggle to handle complex defect types and small-scale targets. This study proposes a novel algorithm based on YOLOv10m for detecting surface defects on smartphone screens. The detection capacity for small and medium-sized objects is improved by replacing the conventional Path Aggregation Network (PANet) with a Bidirectional Feature Pyramid Network (BiFPN). To significantly increase detection accuracy, the original loss function is replaced with the Normalized Wasserstein Distance (NWD) loss function. The enhanced algorithm achieves 59.4% precision, 41.6% recall, and 47.6% mAP50, according to experimental findings. Precision and mAP50 increased by 23.5% and 4.6%, respectively, over the initial YOLOv10m model. Interestingly, these gains were made without adding more parameters to the model, and the detection speed was greatly increased. Additionally, the suggested algorithm’s efficacy in identifying surface flaws on smartphone screens was compared to that of other prominent detection methods. In order to increase detection speed and accuracy, future research will concentrate on further streamlining the network architecture through the use of lightweight networks.