Task-guided dynamic feature selection and loss optimization for multi-scale defect detection

Multi-level features with varying extraction degrees are crucial for detecting multi-scale defects. However, most existing detectors rely on handcrafted heuristics (e.g., IoU) to select positive samples from feature pyramids, which are inherently task-agnostic and unaware of prediction quality. In this paper, we introduce a T ask- G uided dynamic F eature S election (TGFS) strategy, which leverages task-driven classification and localization losses to quantitatively evaluate the matching quality between each ground-truth instance and candidate positive samples across different feature layers during each training epoch. Based on these layer-wise quality scores, each ground-truth dynamically selects positive samples from the first three feature layers that offer the most optimal task-specific representation. To further enhance training effectiveness, we design a non-linear loss reweighting function that adjusts the contribution of the selected samples to the total loss, thereby improving the model’s learning of multi-scale features and enhancing training efficiency. We implement the proposed method on a lightweight, anchor-free FCOS baseline with zero inference overhead. Experimental results demonstrate that our approach achieves competitive performance, reaching 80.7 mAP on the NEU-DET dataset and 70.0 mAP on the GC10-DET dataset.