The work aims to propose a method based on machine vision for surface defect detection of beer metal covers to improve the precision and accuracy of surface defect detection of beer metal covers. Different types of beer metal cover surface defects were used as the research objects. The images were processed by filtering, noise suppression, and high contrast retention algorithms. The YOLO-v5 network was used to complete the defect detection of bottle covers. The original YOLO-v5 model was also optimized by adding the attention mechanism SE module and improving the model loss function and the prediction frame filtering method to suppress the unimportant features in the images and improve the accuracy of small target detection and the feature extraction ability of the model. The comparison between the improved YOLO-v5 model and the commonly used detection model showed that the mPA index of the improved YOLO-v5 model on the test set was 93.10%, and the detection speed reached 294 pieces/min, and the advantages were apparent. For different metal cover surface defects, the detection model based on machine vision has higher detection accuracy and recognition accuracy, and the leakage rate and misdetection rate of minor target defects are low, which meet the real-time and high-precision detection requirements of the production line.