DAF-YOLO for Detecting Small Defects with Weak Visual Features on Chip Surfaces

Abstract

Vision-based inspection of chip surface defects is essential for automated quality control in electronics manufacturing. However, small defects are easily missed, weakly textured defects are difficult to distinguish, and imbalanced class distributions can bias the detector towards majority classes. To address these challenges, this study proposes DAF-YOLO, an enhanced YOLO11 model for detecting small defects with weak visual features on chip surfaces. The model incorporates a dual-pooling channel-spatial attention module (DPCSA) to strengthen salient responses in small-defect regions and suppress interference from complex background textures. An attention-guided detail aggregation module (ADA) recovers edge and texture information lost during deep feature extraction, thereby improving the representation of subtle defect structures. A detail-semantic adaptive fusion module (DSAF) jointly models shallow details, deep semantic cues, and their discrepancies to enhance cross-scale feature fusion. Minority-class focal loss (MCF Loss) is further introduced to increase sensitivity to minority classes and hard-to-classify samples. Experiments were conducted on the D A and D B chip surface defect datasets and the CSDD-det casting surface defect dataset. DAF-YOLO achieved mAP@0.5 values of 86.4% and 79.2% on D A and D B , respectively, exceeding the YOLO11 baseline by 3.6 and 4.2 percentage points. The corresponding mAP@0.5:0.95 values increased by 2.3 and 5.9 percentage points. On CSDD-det, mAP@0.5 improved by 1.1 percentage points. With 3.8 million parameters and an inference speed of 90 frames per second, the model preserves real-time performance while improving detection accuracy. These results demonstrate the effectiveness of DAF-YOLO for chip surface defect detection and its adaptability to different industrial defect scenarios.