The region of interest (ROI) is first extracted under the combination lighting mode of high and low angle light sources and then the median filter kernel of scratch morphology is constructed to obtain the accurate background images from ROI. The scratches are extracted after the background difference. An improved region growing algorithm based on directional gradient is adopted to achieve an effective connectivity for the same scratch, which reduces the miss rate of scratch detection. By analyzing the confidence of main scratch detection parameters such as area, length-width ratio and so on, a scratch detection method based on weighted fusion of multi-features is proposed. The results show that, for this method, the accuracy of scratch detection is 95.7%, and the processing time is less than 1.21 s, which meets the accuracy and efficiency requirements for the engineering application.