2.5 ROI extraction

The extraction of ROI is shown in Algorithm 1, as illustrated in Fig 13.

Algorithm 1: ROI Extraction Process

Requirements: Hand image I, object detection network YOLO5-lite, and keypoints detection network Improved UNet

1:Input I to YOLO5-lite to obtain the palm localization image I1.

2:Input I1 to Improved UNet to detect two keypoints, A and B, where A represents the valley point between the index and middle fingers, and B represents the valley point between the ring finger and little finger.

3:Calculate the angle θ between the line AB and the horizontal direction, θ = arctan (|yB−yA||xB−xA|) and rotate the image by θ to correct its orientation.

4:Establish a Cartesian coordinate system with the line AB as the x-axis and the perpendicular bisector of AB as the y-axis, with the origin at O.

5:Set point C as the point on the y-axis, with a distance of 0.8|AB| from O.

6:Draw a square with point C as the center and a side length of 1.2|AB|, and extract it as the ROI of the palm print.

It should be noted that the selection of midpoint C and the setting of ROI edge length are flexible and variable, and may not necessarily be 0.8|AB| and 1.2|AB| in practical applications. This is also one of the advantages of this method, which is to determine the position of key points and flexibly set the ROI region to be intercepted based on different distance rules, enabling multi-scale ROI region extraction.