Weighted post-processing method applied to face detection prediction box

Abstract

The invention relates to the field of face detection, and discloses a weighted post-processing method applied to a face detection prediction box, which comprises the following steps: performing face detection on a picture to be detected by using a face detector; deleting the prediction box of which the confidence is lower than a first confidence threshold; calculating a generalized cross-parallelratio between the prediction box with the highest current confidence degree and the updated prediction boxes; screening out n prediction boxes of which the generalized intersection-combination ratiosare greater than an intersection-combination ratio threshold value from the updated prediction boxes; obtaining a weighted prediction box corresponding to the prediction box with the highest current confidence; deleting the n prediction boxes; and obtaining a weighted post-processing result of the face detector. According to the invention, effective prediction data of more face detectors are utilized, so that the position information and confidence information of the deleted prediction box are effectively utilized at the same time, the position of the output prediction box is better corrected,a generalized cross-parallel ratio is introduced, and the position precision of final output of the face detectors is improved.

Description

technical field [0001] The invention relates to the field of face detection, in particular to a weighted post-processing method applied to a face detection prediction frame. Background technique [0002] With the wide application of deep learning technology, face recognition, face beauty and other technologies are becoming more and more mature, and face detection is the first step in these application scenarios, and its accuracy has a greater impact on the accuracy of other subsequent algorithms. Impact. In the face detection algorithm, the model will predict many coordinates and corresponding confidence levels that may be face frames. These prediction frames often cannot be used as the final output. It is necessary to combine repeated prediction frames and predictions with lower confidence. box to filter out. [0003] Most of the existing methods use the non-maximum suppression (NMS) method. First, the intersection ratio (IoU) of the prediction frame with the highest conf...

AI Technical Summary

Problems solved by technology

This calculation method only considers the prediction frame with the highest confidence in the current picture, and deletes other prediction frames that have a high degree of coincidence with the IoU calculation. These deleted prediction frames also contain the location information of the face, resulting in more loss of valid data

Moreover, when IoU is used as the calculation method of the coincidence degree, it cannot distinguish the different alignment methods between the two prediction frames, and cannot accurately reflect the degree of overlap of the two prediction frames.

In the final output result of this method, only one prediction frame information is used for a face, and a lot of valid data is not used, and the position accuracy of the final output face frame is not high.

Images