DCAFD-YOLOv8: Dynamic Class-Aware Focal Feature Distillation for Class-Imbalanced Object 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).

In industrial visual detection, class imbalance often causes model optimization to favor high-frequency classes, leaving low-frequency targets insufficiently represented and more susceptible to missed detections. At the same time, lightweight model design further constrains feature representation capacity. To mitigate the representation loss introduced by model compression, a frozen teacher network is used to transfer multiscale knowledge from the P3, P4, and P5 feature levels to the Light-YOLOv8-base student model. To reduce background interference during conventional feature distillation, a foreground-focused mechanism is introduced to concentrate knowledge transfer on informative target regions. A class-frequency prior is further incorporated to compensate for insufficient supervision of low-frequency classes. Since fixed weighting cannot adequately reflect the evolving learning states of different classes, class-wise teacher-student feature discrepancies are used to dynamically estimate learning difficulty and adjust the distillation strength accordingly. In addition, teacher confidence is introduced to suppress unreliable supervisory signals. These components together form DCAFD-YOLOv8 and are activated only during training, leaving the student network unchanged at inference. Experimental results show that the proposed model requires only 0.9 M parameters and 2.4 GFLOPs, while achieving an overall Recall of 90.6% and an mAP50 of 94.5%. The Recall of the low-frequency coal class improves from 90.6% to 91.2%, indicating that DCAFD-YOLOv8 can improve low-frequency target detection while preserving the advantages of lightweight deployment.