MemMat-YOLOX: An Intelligent Detection Method for Weld defects for DR Video Images

DR digital real-time imaging technology can generate continuous video data and has become an important method for detecting large diameter spiral pipe welds. However, the existing manual viewing of DR video data recognition defects can easily lead to poor consistency and low efficiency in the results. Meanwhile, current defect detection models lack the ability to integrate temporal information, which can easily lead to the lost of small scale defects information in DR video. To solve the above problems, a defect detection method based on MemMat YOLOX for DR video data of pipe welds has been proposed. First, the existing multi-scale convolution structure is improved by building matrix network module (MatrixNets), and the detection ability of large aspect ratio features of defects is improved. Then, a diversity perception feature memory aggregation module (DP-FMAM) is introduced, which is able to store and update diversity feature information in the video data. Finally, for the problem of weak feature expression of single-frame DR image, multi-head attention mechanism is integrated to improve the ability of detecting the model. Through the verification on the actual enterprise data set, the results show that the proposed method improves the video defects precision and reduces the lost.