YOLO-Xray: A Bubble Defect Detection Algorithm for Chip X-ray Images Based on Improved YOLOv5

• DOI: 10.3390/electronics12143060
• Publication Date: 2023-07-13T05:52:25Z
• Authors: Jie Wang; Bin Lin; Gaomin Li; Yuezheng Zhou; Lijun Zhong; Xuan Li; Xiaohu Zhang
• Keywords: Backbone network; Feature (linguistics)

In the manufacturing of chips, accurate and effective detection internal bubble defects chips is essential to maintain product reliability. general, inspection performed manually by viewing X-ray images, which time-consuming less reliable. To solve above problems, an improved defect model YOLO-Xray based on YOLOv5 algorithm for chip images proposed. First, are preprocessed image segmentation construct dataset, namely, CXray. Then, in input stage, K-means++ used re-cluster CXray dataset generate anchors suitable our dataset. backbone network, a micro-scale head added improve capabilities small detection. neck bi-direction feature fusion idea BiFPN new network fuse semantic features different layers. addition, Quality Focal Loss function replace cross-entropy loss imbalance positive negative samples. The experimental results show that mean average precision (mAP) reaches 93.5%, 5.1% higher than original YOLOv5. Meanwhile, achieves state-of-the-art accuracy speed compared with other mainstream object models. This shows proposed can provide technical support images. also open available at