SED-YOLO: an enhanced YOLOv8 algorithm for micro defect detection in polariser films

Polariser films are indispensable components of LCD panels; however, they are susceptible to defects like stains and scratches during the manufacturing process. These flaws can compromise the imaging quality, causing significant economic losses. Traditional manual inspection methods suffer from inefficiency and high subjectivity. To address this, we present SED-YOLO, an enhanced YOLOv8nmodel designed for the detection of micro-defects in polariser films. SED-YOLO integrates four key innovations: Space-to-Depth Convolution (SPD-Conv), a cross-stage partial bottleneck with two convolutions and Efficient Multi-scale Attention (C2f-EMA), a dedicated small-object detection branch with a Dynamic Head (DyHead), and the Wise Intersection over Union (WIoU) loss function. These innovations strengthen the model’s capability by preserving critical details, reweighting feature representations, enhancing cross-scale information integration, and optimising bounding box localisation. On a benchmark polariser defect dataset, SED-YOLO achieves 91.6% precision, 89.8% recall, and 93.8% mAP@0.5, representing improvements of 2.3%, 1.8%, and 2.4% over standard YOLOv8n, respectively. Running at 344 FPS on an NVIDIA GeForce RTX 3080 Ti, the model maintains real-time performance, offering a reliable framework for high-precision industrial inspection and quantitative non-destructive evaluation (QNDE).