X-ray weld image information detection based on lightweight YOLO

To accurately identify the identification marks containing a lot of information on X-ray weld images,the YOLO-G and YOLO-D were proposed respectively.YOLO (you only look once)-based X-ray weld image information detection method was designed,and the number of model parameters and the amount of calculations were reduced significantly through the introduction of a variety of lightweight technology,achieving a high detection accuracy while speeding up the detection speed and reducing the dependence on high power hardware.Meanwhile,the prediction layers of network and the size and number of anchor were adjusted adaptively according to the characteristics of the detected objects.Through the experiment on the data set,results show that compared with YOLO-V3,YOLO-G and YOLO-D achieve higher accuracy,and the model sizes are reduced by 89.6% and 79.68% respectively,with the computational cost of only 11.4% and 21.3% of YOLO-V3,respectively. © 2021, Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.