Research On Information Detection Method Of X-ray Weld Image Based On Lightweight Neural Network

The establishment of X-ray image database can improve the service level of film management and X-ray inspection.When the special equipment enterprises establish the information base,they need to detect the weld information in the image and take it as the warehousing basis.The information detection process can be regarded as the object detection task in machine vision.In recent years,convolution neural network has achieved good results in the field of object detection.In order to achieve higher detection accuracy,network design is developing in the direction of deeper layers and more complex structure.However,the huge amount of calculation and parameter leads to higher requirements on the hardware calculation and storage capacity,which makes deep learning unable to be well applied in the practical application with limited equipment resources.In view of the above problems,this paper proposes a X-ray weld image information detection method based on lightweight neural network according to the characteristics of weld image and its information characteristics.This paper optimizes YOLO-V3 and tinyyolo network respectively based on difference of detection task difficulty.And the specific improvements are as follows:(1)K-means + + is used to cluster anchor box size based on weld image data,making anchor box size suitable for the task of this paper;(2)group convolution and depth separation convolution are introduced to achieve network lightweight;(3)the network structure is adjusted according to the characteristics of the tested object,and the yolo-z and yolo-s networks are designed and proposed for center mark and identification mark detection,respectively.At the same time,aiming at the situation that the brightness of different weld image varies greatly,this paper proves that the convolution neural network has strong robustness to the change of illumination through the visual analysis of intermediate feature maps,which can effectively solve the problem that the difference of illumination brightness affects the detection accuracy of weld image information.Finally,dataset 1 and 2 are introduced to compare different networks in terms of their precision,recall,ect.The experiments show that the proposed yolo-z and yolo-s can greatly reduce the size of the model,reduce the computational complexity,accelerate the model training and detection speed,and reduce the requirements of carrying network equipment while ensuring high information detection accuracy.The proposed methods canreduce the economic cost and time cost of the enterprises,which is very practical and real time.