Weld Defect Classification in Radiographs Using HOG, SVM and CNN

Crack recall of the deployed CNN on unseen welds is only 0.093 (95 % CI 0.007-0.264): this is the safety-relevant result of this project. Radiographic testing produces films that qualified interpreters read by eye. Using the public RIAWELC patches (24,407 grey patches from 29 welds: crack, porosity, lack of penetration, no defect) this project compares CLAHE + HOG + LBP + SVM with a small CNN and audits the released folders first: the testing folder is 2,443 of 2,443 byte-identical copies of training files, and every test patch's weld is also in training. On that patch protocol the SVM reaches 0.767 and the CNN 0.744; on welds never seen in training the deployed CNN reaches 0.530 and the SVM 0.517, with no significant difference. The specification's targets are not met. PR5 is the only project of the group with X-ray weld radiographs (unlike A13 steel surface images, E12 PCB template pairs, C1 concrete photographs and E11 wafer maps); its central issue is weld-level leakage and crack recall. A FastAPI web application with five interchangeable databases keeps the decision with the film interpreter. Research and teaching prototype, not a certified radiographic interpretation system: film reading and the acceptance decision stay with a qualified RT level 2 / 3 interpreter.

Category: Production Engineering

Production Engineering Project PR05 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.