Abstract: Aiming at the problems of low contrast of industrial X-ray weld images, fuzzy defects and relatively small areas, and they were difficult to identify, a recognition framework combined with convolutional neural networks was designed.According to the characteristics of the defect image, the size of the corresponding neural network structure, convolution template and pooling template were designed.Based on the analysis and determination of the neural network structure, the sensitivity and training algorithm of the convolutional neural network were also given in the article.The effectiveness of the neural network structure was verified through examples, the defect detection accuracy rate was 97% and the false alarm rate was only 3%.At the same time, the X-ray weld image suitable for the recognition of the convolutional neural network was analyzed, it was found that the convolutional neural network with the effective information span of the gray histogram above 50 could be effectively identified.It showed that the neural network designed in this paper was feasible and effective in the recognition of X-ray weld defect images.Highlights:(1)The convolutional neural network structure was set to 4 level 6 layers, convolution kernel size was 3 ×3, the depth of the original image was 1.After first convolution operation, there were four channels.After 3×3 convolution kernels convolution, there were three channels to identify defects, and suspected defect area could be preliminarily confirmed.After the second convolution operation, there were 16 channels, and then after the convolution of 3×3 convolution kernel again, the defect image could be identified.(2)By analyzing the gray histogram of successfully recognized and failed recognized images, when the gray scale span was less than 50, the contrast between the image detection target and the background was relatively lower, and the convolution layer in the convolutional neural network was difficult to extract the target features.In this case, the recognition of the neural network fails and defects could not be determined.When the gray histogram span of the local image was higher than 50, the recognition target was more prominent than the background, the overall image contrast was higher, and the convolution layer could extract the features of the recognition target.In this case, the convolutional neural network had a good recognition effect on the weld image.