Extracting region of interest for palmprint by convolutional neural networks

Palm ROI extraction is one of the most important processes in palmprint recognition. The core idea is to employ the valley points between the fingers to establish a coordinate system and then obtain the ROI of palmprints. However when extracting the keypoints, conventional methods have three problems: (i) they are so sensitive to parameters and background noise due to lack of joint optimization, (ii) accuracy of the location of keypoints is not good enough, (iii) extracting speed can be faster. To address the above problems, this paper presents a novel approach to extract palmprint ROI using convolutional neural net. First, we present a new CNN to identify the palmprint being a left or right hand. Then we propose a specific designed and optimized CNN to detect the keypoints. Finally we test our method using palmprint verification algorithm, competitive coding. Experimental results show that the proposed novel method is not only fast and efficient, but also robust for ROI extraction.