Surface defect detection on printed playing cards using an improved YOLOv8n model

Abstract Reliable online inspection of playing card surfaces is difficult because small, low-contrast defects are easily confused with dense printed graphics. This study proposes a defect-aware YOLOv8n framework integrating a ResNet50 backbone, DeCA, an improved BiFPN with SPD-Conv, a dynamic adaptive decoupled head, and Wise-IoU v3. These components strengthen weak-feature extraction, suppress background interference, preserve multiscale details, and improve localization. A production dataset of 2750 original images covering edge defect, ink missing, fold, and stain was augmented to 6000 images and divided 7:1:2 at the source-image level to prevent information leakage. Across three independent training runs, the proposed model achieved $$96.27\pm 0.16\%$$ mAP@50 and $$77.07\pm 0.18\%$$ mAP@50:95, corresponding to mean improvements of 12.77 and 11.64 percentage points over YOLOv8n, respectively. It also reached 97 FPS with 21.85 M parameters on the experimental GPU, indicating a favorable detector-level accuracy-speed balance under the evaluated hardware conditions. The evaluation is restricted to one production-line dataset; no external or cross-domain validation was conducted, and no generalization or transferability claim is made.