Defect Detection Algorithm for TFT-LCD Circuits Based on Modified YOLOX Model

Detect TFT-LCD circuit defects is a known challenge in the industry. To take on this challenge, a modified YOLOX defect detection model is proposed to improve accuracy and the speed of multi target detection. Under proposal, convolutional block attention module (CBAM) is introduced into the Backbone and the Neck of the model to reduce background interference. Also, efficient task-aligned Head (ET-Head) is employed at the inference module to tackle the problem of independent classification and localization tasks. And task alignment learning (TAL) is used to link these two tasks. Further, varifocal loss is modified for classification task to balance category differences; SIoU loss function and distribution focal loss are applied for localization task to optimize bounding box regression. The effectiveness of the developed method is verified through experiments where mean Average Precision (mAP) increased by 3.69% against the benchmark model (YOLOX). The developed method also outperforms other mainstream object detection models in terms of detection accuracy in the experiments and it shows great application potential in the TFT-LCD manufacturing industry.