Abstract: The defect detection using manual method requires a lot of manpower and high cost.Moreover,visual fatigue and differences in human subjective understanding are easy to produce judgment errors.The traditional method of defect detection has poor anti-interference ability,complex parameter setting,high application threshold,and poor robustness to complex background and defects.In order to solve the problem,an integrated software system is developed,which can easily and quickly build models and evaluate the performance of deep learning algorithms.The system has the functions of labeling,training,detection,review,and integrates the target detection algorithm with the functions of end-to-end defect recognition,target location and classification,which can achieve pixel-level segmentation and classification of defects.Meanwhile,the system can classify and detect normal and abnormal samples only by training normal samples,which has great application potential.