YOLOV7-TMV3: Lightweight Non-Contact Palm Print Region of Interest Segmentation Model

With the development of science and the progress of the times, more and more biometric technologies including face recognition, fingerprint recognition, palmprint recognition, iris recognition, human posture recognition, etc. have great development potential in the fields of people's clothing, food, housing and transportation. Palmprint recognition has the characteristics of immutability and weak invasiveness. The extraction of stable, robust and complex environment segmented palmprint region of interest is one of the key problems of palmprint recognition. Existing methods are sensitive to noise, light and dark changes, and mostly perform palmprint region of interest extraction in a single background, so unconstrained and complex background palmprint region of interest extraction for smart devices is relatively rare. We propose a lightweight palmprint ROI keypoint detection model YOLOV7-TMV3 based on yoloV7- tiny improvement, we combine the target detection with the algorithm without separating the palm from the background, and establish the palm segmentation coordinate system by keypoints to segment the palmprint region of interest according to the scale. The feature extraction backbone of YOLOV7-TMV3 is MobileNetV3. The improved model reduces the number of parameters by 33%, increases the MAP by 10
^%
The number of floating point operations is reduced from 13.2 to 6.9 GFLOPs, a 47% reduction, which greatly reduces the hardware requirements and has the possibility of application on smart devices, which has a certain practical value.