Improved Welding Defects Recognition with Transfer Learning Based You Only Look Once Version 5

In recent years, welding is a dynamic joining procedure for infrastructure development and energy production to ensure safety and reliability. However, weld defects often reduce the quality of welded parts. Existing TL-CNN models struggled with variations in defect appearance, leading to biased training and poor generalization. To overcome this issue, TL based You Only Look Once version 5 (TL-YOLOv5) is proposed. The pretrained YOLOv5 model is fine tuned for optimizing the welding data to generalize varied appearances. Initially, the input data is collected from asymmetric laser weld images and then preprocessed with median filter for image denoising and then enhanced with gray scale transformation. After that, Convolutional Neural Network (CNN) with improved activation function and adaptive pooling is incorporated to extract features from the preprocessed images. Finally, the input feature maps are further fed into the proposed TL-YOLOv5 for weld defect recognition with anchor boxes and fine-tuned to improve performance. From the results, the proposed TL-YOLOv5 attained better results in contrast to conventional TL-CNN including accuracy (99.12%) respectively.