GDTS-YOLOv8: Gather-and-Distribute Mechanism With Two-Stage Self-Fine-Tuning Strategy for Multiscale Capacitor Defect Detection Under Imbalanced Data

Capacitors are essential in industrial applications due to their ability to store and release electrical energy rapidly. However, capacitors, with their complex manufacturing processes and heterogeneous characteristics, pose two critical challenges to conventional detection algorithms: 1) the difficulty in detecting small-scale defects amid multi-scale defect variations; and 2) the inherently imbalanced distribution of defects. To address these issues, the GDTS-YOLOv8 detection framework is introduced, which leverages a Gather-and-distribute mechanism (GD) and a two-stage self-finetuning strategy with a frozen backbone (TS). Firstly, GD is constructed to enhance the efficiency of multi-scale feature extraction and fusion in the YOLOv8 feature fusion layer. Centered on optimizing cross-level information exchange and reducing information loss, this mechanism can effectively address the issue of poor detection performance for small targets in multi-scale scenarios. Secondly, TS is proposed to address the inherent imbalance problem, comprising: (a) a rare-defect upweighting pre-training stage with label reduction of frequent defects; (b) a full-data fine-tuning stage integrating label reintroduction of frequent defects with backbone freezing for optimal performance. Experimental results confirm the proposed method's effectiveness, enhancing capacitor detection precision by over 5% (75.3% peak accuracy) and demonstrating significant improvements in small-defect detection and imbalanced sample handling.