KDSP-DEA-YOLO: Knowledge-Distillation-Guided Structured Pruning for Lightweight Industrial Surface Defect Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Small surface defects on industrial components are characterized by small scale, low contrast, weak texture, and strong background interference. To improve deployment efficiency while preserving detection accuracy, this study proposes KDSP-DEA-YOLO based on the DEA-YOLOv8n teacher model. A heterogeneous lightweight student is first constructed using GhostConv, C2f_Ghost_Lite, a lightweight EMA branch, and DFF_Lite, reducing the pre-pruning parameter count to 4.32 M. A three-stage pipeline is then adopted: sparsity-aware pre-distillation, dependency-aware structured pruning, and distillation-assisted fine-tuning, with joint supervision from outputs, multiscale features, and bounding boxes. Based on the current baseline and compression trends, the predicted 20% pruning configuration contains 3.61 M parameters and requires 8.84 G FLOPs, while achieving approximately 69.8% mAP@0.5 after fine-tuning. This represents a 39.8% reduction in parameters relative to the 6.0 M teacher. The framework provides a verifiable route toward regularized compression and edge deployment for small industrial defect detection.