TAFF-YOLO: An Enhanced YOLO Model With Transformer and Attention Feature Fusion for Chip Internal Defect Detection

Abstract

Accurate and efficient detection of internal defects in chips is crucial for ensuring the reliability and yield of electronic products. However, conventional object detection models often struggle to identify such defects in ultrasonic images due to their extremely small size and low contrast. To overcome these challenges, a novel detection framework that combines image processing with a Transformer-based Attention Feature Fusion YOLO model (TAFF-YOLO) is proposed in this paper. The framework first employs image processing methods to extract individual chip images and then applies a handcrafted feature generation algorithm for data augmentation to mitigate overfitting. Subsequently, the TAFF-YOLO model is introduced for precise defect localization. Specifically, a Deformable Attention Background Suppression Module (DABSM) is integrated to enhance global contextual information and the receptive field for detecting small objects. Additionally, the feature fusion network (neck) is augmented with a ResBlock and Convolutional Block Attention Module (ResCBAM) to focus the model effectively on critical defect regions while suppressing irrelevant background noise. A Smooth Intersection over Union (SIoU) loss function further stabilizes the training and enhances localization accuracy. Experimental results from two private datasets and one public dataset demonstrate that the proposed method achieves an average precision of 90.9%, with a compact model size of 2.6 million parameters and an inference speed of 167.3 FPS, effectively balancing detection accuracy, model complexity, and processing speed. This indicates significant promise for real-time industrial chip inspection applications.