Abstract: In order to improve the detection rate, recall rate and detection efficiency of steel welding defects, a steel welding defect detection method based on machine vision and convolutional neural network was proposed. Graying processing, mean filtering and median filtering methods were used to enhance the image; welding defect features were extracted based on robust principal component analysis（RPCA）, and then the defects were localized by threshold segmentation and morphological corrosion-expansion operations. For the welding defect detection problem, a classifier was built by improving the traditional ResNet152 convolutional neural network model for training and it was compared with the traditional ResNet152, InceptionNet-v3, ResNet101, DenseNet and ResNet50 classifiers. The test results show that the comprehensive performance of the improved ResNet152 classifier is better than those of the other five network models, the recall rate is 99.05%, the checking accuracy rate is 98.11%, the mean average precision is 97.54%. The present method has a certain degree of practical significance for the detection of welding defects of steel.