Weld Image Evaluation Model Using Improved Loss Function

This paper presents a study focused on the problem of evaluating the quality of weld seams in the context of metalworking practice. The main challenges encountered in this task are the high similarity between weld images and the presence of inconspicuous grading boundaries. To address these challenges, we propose a novel loss function that incorporates the VGG16-SVM and Mask R-CNN models, incorporating multiple classifications and introducing central loss and isolation loss terms. To construct a comprehensive sample dataset, various image data types were acquired, followed by data enhancement, size modification, normalization, and other preprocessing operations on the weld images. Finally, practical validation experiments were conducted using two different models, in combination with the improved loss function model. The experimental results provide compelling evidence for the effectiveness of the proposed method.