Extracting palmprint ROI from videos with complex backgrounds based on hand pose estimation

Current research on the extraction of palmprint region of interest (ROI) predominantly focuses on fixed scenarios and static images, with relatively fewer studies addressing the extraction of palmprint ROI from videos against complex backgrounds. It is imperative to address the challenge of extracting palmprint ROI from videos with complex backgrounds to enable the broader application of palmprint recognition technology in practical scenarios. We propose a novel ROI extraction method for videos with complex backgrounds to tackle this issue. Our approach utilizes a hand posture estimation model to filter out images in dynamic videos that do not meet the requirements for palmprint ROI extraction and those with poor image quality. Subsequently, we employ semantic segmentation techniques to extract high-precision interdigital contour information from images containing complex palmprint backgrounds, thereby obtaining the palmprint ROI. Compared to other methods, this approach demonstrates better adaptability to various complex scenarios. We have also collected a database (HFUT Video), comprising 400 videos, specifically for the extraction of palmprint ROI from videos with complex backgrounds. Experimental results on the HFUT Video database indicate that the proposed method achieves a 100% accuracy rate in localization and extraction palmprint ROI on the database, exhibiting excellent performance in videos with complex backgrounds and holding great promise for practical applications.