Welding Defect Inspection Using Deep Learning

To supplant human shortcomings in Welding Inspection System performed manually by humans, an intelligent computer vision inspection leveraging deep learning technologies method is proposed. In the proposed model, the top layer of the VGG16 model was replaced by a fully connected layer with a softmax activation function. Rather than providing an extensive comparison between different transfer learning models (Inception V3, Xception, VGG16, VGG19, ResNet50), we analyzed the VGG16 model in-depth, as it performed well in initial comparison on accuracy and training time as compared with other models. All models were trained using a welding image dataset to identify and detect various welding defects, including Cracks, Over-Roll, Under-Fill, Porosity, and Mechanical Damage. The proposed VGG16 model was evaluated to verify how accurately, faster, and precisely it performs multiple-class classification. Furthermore, the proposed model was optimized by hyper-parameter tuning of Learning rate and Batch size, and an overall 8.5% increase in the accuracy was observed. The experimental results show that in several aspects, VGG16 had performed quite accurately in classifying defects and achieved the highest accuracy of 85%.