An X-ray imaging weld detection method based on depth learning

Abstract

The invention relates to an X-ray imaging weld detection method based on depth learning. The method comprises the following steps: step S1, screening and arranging the original X-ray imaging weld digital images, labeling the images on the weld area, and completing the production of a sample image set; S2, finely tuning the sample image set by using the ResNet 101 network pre-trained on the ImageNet to complete the initialization of the model; S3, respectively training an RPN convolution neural network and a fast RCNN convolution neural network for generating candidate regions by using the prepared sample image set to obtain a trained detection model; S4: Input the picture to be detected into the trained detection model and output the detection result. In accordance with that prior art. Theinvention fully utilizes the high-efficiency feature extraction ability of the ResNet 101 basic network and the high-precision advantage of the Faster RCNN convolution neural network for image recognition, avoids misjudgment and omission caused by manual evaluation subjectivity, and has the advantages of high robustness, high accuracy and low omission detection rate.

Description

Technical field [0001] The present invention relates to a welding seam detection method, and in particular to an X-ray imaging welding seam detection method based on deep learning. Background technique [0002] Computer vision is an important interdisciplinary subject in the fields of artificial intelligence and image processing. Early methods for solving computer vision tasks mainly included two steps, one was to manually design features, and the other was to build a shallow learning system. With the development of artificial intelligence, deep learning was officially proposed in 2006. Computer vision was the first field where deep learning technology achieved breakthrough achievements. Target detection is a very important research hotspot and direction in the field of computer vision, involving many disciplines such as image processing, machine learning, pattern recognition, etc. Its ultimate goal is to simulate human visual ability so that computers can be as fast as hum...

AI Technical Summary

Problems solved by technology

This method needs to construct a multi-scale filter operator according to the optimization parameters of different images, the scope of application is limited, it is not universal, and its robustness is not high

[0007] Chinese Patent Document No. CN103914838A Publication (Announcement) Day 2014.07.09 discloses a method for identifying defects in industrial X-ray weld images. First, the sample image is set, and the size is normalized uniformly. The pattern recognition problem of the image, Converting it to an equation system to solve the problem, this method needs to determine the number and size of sample images in advance, and cannot detect weld images of any size. check

Images