Rotating palm image detection method

Abstract

The invention discloses a rotating palm image detection method, which comprises the following steps of: 1, rotating an image to be detected at an interval of 5 degrees to form a 72-original image, andstoring the 72-original image into an image rotating star disk from a rotating image within a range of 0-360 degrees; 2, inputting all the pictures in the image rotation star disk into a pre-trainedforward palm detector in batches according to a sequence; 3, screening all detection results to the maximum extent, and obtaining a unique candidate detection result; 4, calculating a reverse rotationmatrix corresponding to the rotation angle in the candidate detection result; and finally, performing rotation processing on the candidate detection result through the rotation matrix to obtain a final palm image detection result with an angle. The invention provides an image rotation star disc and a maximum possibility screening method, and combines the most efficient detection network Yolov3 inthe target detection field in the current deep learning to realize palm image detection with an attached angle.

Description

Technical field[0001]The invention relates to the technical field of image detection and image recognition, in particular to a method for detecting palm images at any angle combined with a fusion image rotating astrolabe, maximum possible screening and a forward palm detector.Background technique[0002]Palmprint images have rich texture information and can be multi-modally fused with fingerprint features to further improve the accuracy of recognition. Therefore, more and more scholars in the field of biometric recognition have begun to devote themselves to this research. The extraction of regions of interest plays a vital role in palmprint recognition. Traditional palmprint recognition algorithms use coordinate transformation to extract regions of interest. This method relies on the detection of key points. The requirements for palm images are relatively high. Well used in unrestricted environments. With the continuous maturity of related technologies in the field of target detection...

AI Technical Summary

Problems solved by technology

However, the traditional target detection general algorithm [1, 2, 3, 4] focuses on the detection of the target position and size and ignores its angle. Therefore, when using this algorithm for palm image detection, only the approximate area where the palm is located can be detected. It will have a great impact on palmprint recognition, because the palmprint information of the palm changes in different directions. Only the palm with the rotation angle detected can be obtained in the subsequent region of interest extraction with the same direction. sexual palm area

[0003] For the detection of rotation angle, there have been many studies on face recognition. One is the method of data enhancement, that is, adding faces of various angles in the training set for training, but this idea greatly depends on the diversity of samples. , at the same time, in order for the network to learn so much information at the same time, a larger network structure is also required, and the detection results cannot show the specific rotation angle, and can only output the approximate area where the position of a face is located.

H.A.Rowley[5] and others proposed a Router face detection network structure as early as 1998. The network structure detects the face, calculates the angle of the face, and then rotates the face according to the angle, and then only trains one The classifier for the frontal face is fine, but the disadvantage is that the angle is a 360° regression problem, and the error tolerance space is too large to obtain a sufficiently accurate angle prediction.

In 2007, C.Huang[6] and others proposed to use face samples from different angles to train multiple detectors. The disadvantage is that they are flat, and the angle of faces in the plane is 360°. To train so many classifiers, the time complexity and Model complexity is too large

This method is very time-consuming because it needs to train four networks, and because PCN-3 directly regresses the angle, the angle error in the face data reaches 8°, which is very important for the extraction of the region of interest in palmprint recognition. unacceptable error

Images