Welding Defect Detection Using Deep Learning and Semantic Segmentation

Numerous computer-aided techniques have been created to automatically detect defects and imperfections in welding radiography images. The purpose of these strategies is to circumvent the challenges that are associated with human interpretation. In the subject of welding problem diagnostics, deep learning algorithms have garnered a substantial amount of attention due to the inherent benefits that they deliver. These advantages include powerful generalization capabilities and automated feature extraction techniques. Additional research is required, however, to investigate several various areas. These features include the gathering and improvement of welding radiography data, the selection and development of deep neural networks, and the enhancement of the model's capacity to be applied to a variety of situations while being easily understandable. This work intends to construct an automated system that can identify defects in welding by utilizing a semantic segmentation technique on radiographic images taken from a specific dataset of welds. This is taking into consideration the context in which this study is being conducted. Utilizing the UNet architecture, the procedure entails the training of several models, in addition to the incorporation of transfer learning strategies that utilize multiple backbones. A comprehensive comparative analysis is carried out after the model has been developed, and the evaluation criteria that are used are those that have been developed expressly for deep learning models. The results of the tests demonstrate that the method that was recommended is effective in precisely defining the boundaries of the various types of Defects and in faithfully identifying the various types of defects. In addition, the findings demonstrate that the system is capable of doing an automatic analysis of radiographic images to detect welding defects. Due to this fact, it is an indispensable component for quality inspection systems that are utilized in welding tasks.