This study proposes a video image time watermark detection method based on PaddleOCR, which addresses the issue of manual frame-by-frame reading of time watermarks in vehicle speed judicial appraisal practices. This study collects and labels 20 video materials containing time watermarks, which are classified into four types: pure black, pure white, black-and-white interlaced, and black edges filled with white, totaling over 15,000 frames. The images cover typical scene types in judicial appraisal practice, such as daytime, nighttime, rainy, foggy, sunny scenes, low-resolution CCTV, and compressd videos conditions. A four-stage detection model was designed, and the automatic detection of time watermarks in video images was achieved through a pipeline: PaddleOCR for single-image detection, image processing, PaddleOCR for video recognition and data cleaning with post-processing. Experiments demonstrate that the method proposed in this paper achieves an average accuracy rate of about 98% in the dataset and can be applied to identification practice.