Welding defect detection is an important technical method and means to ensure welding quality in the welding process. The traditional manual inspection has the problems of low accuracy and low measurement efficiency, which cannot meet the requirements of modern industrial production. With the continuous development of artificial intelligence, using machine vision instead of human to detect welding defects intelligently has become the direction of development. The main processes of welding defect detection based on machine vision is to first collect images, then analyze and preprocess images, and finally use artificial intelligence to identify and classify welding defects. Therefore, this paper mainly summarizes and analyzes the current research progress of welding defect detection from two aspects: welding image acquisition and preprocessing, and defect detection and classification methods. It focuses on the defect detection and classification methods based on machine learning. Furthermore, this paper puts forward the problems to be improved and solved, and looks forward to the future development trend.