Aiming at the four major defects such as scratches, missing points, stones and watermarks on the glass surface, the work aims to study a method for detecting surface defects of glass based on backlight illumination, and propose an improved K-means clustering algorithm for watermark defect detection. First, the glass image was acquired by the designed image acquisition system, and the background of the acquired image was estimated. Then, based on the gray difference between the watermark defect and the other three types of defects, the defective glass was divided into two categories to complete the rough classification of the defects. Next, the edge detection algorithm was used to process the glass image with such defects as scratches, missing points and stones, and the improved K-means clustering algorithm combining Otsu threshold segmentation method and compensation coefficient f was used to process the glass image with watermarks. Finally, the identification and marking of four defects on the glass surface were completed. The experiments showed that the system was easy to operate, the algorithm was low in complexity, the accuracy of defect identification was high, and the detection speed was fast. Through the above-mentioned glass surface defect detection system, four major defects on the glass surface can be accurately and efficiently detected. The improved K-means clustering can accurately detect watermarks. It overcomes the shortcomings such as a huge number of cluster iterations and clustering results easy to fall into the local minimum. As the proposed method greatly improves the efficiency of defect detection, it can be used for the real-time detection in the process of glass production.