A real-time robust face detection method

Abstract

The invention relates to a computer vision recognition technology, in particular to a real-time robust face detection method. The method comprises the steps of obtaining a target image needing the face detection and processing; establishing and training detection model, wherein the detection model includes a plurality of convolution module, multiple Inception modules, a plurality of Inception modules with residuals and a plurality of detection modules, inception module is a channel separation and convolution module with two branches. Inception module with residual error is a channel separationand convolution module with residual error connection, and the detection module uses convolution operation to output position information and classification information; inputting the target image into the trained detection model and obtaining the convolution results at the specified level; carrying out the classification and regression of the convolution results obtained; according to the results of regression and classification, calculating the position of human face. The method constructs a simple and efficient convolution neural network, reduces the redundant operation in the detection process, and achieves real-time effect on the CPU.

Description

technical field [0001] The invention relates to computer vision recognition technology, in particular to a real-time robust face detection method. Background technique [0002] Face detection is a sub-topic of target detection in computer vision. It belongs to a specific category of target detection. It is a classic problem that has been deeply studied in machine vision. It has important application value. The goal of face detection is to find the location of a face in an image. Since Viola and Jones proposed the use of Haar-like features and cascaded AdaBoost classifiers to detect faces in 2001, new features (such as LBP, Bow, HOG, etc.) and new detectors (such as SVM, LatentSVM, DPM )Been proposed. Among them, the DPM (Deformable Part Model) algorithm is a component-based detection algorithm, which has a good detection effect on distorted, multi-pose, and multi-angle faces. [0003] At present, detection algorithms based on artificially designed features are easily aff...

AI Technical Summary

Problems solved by technology

[0007] The detection speed of the Cascade CNN method is greatly affected by the number of targets to be detected. The larger the number of targets, the more regions need to be screened, and the slower the detection speed; the two-stage method has more suggested regions to extract. The detection speed is generally slow; hyperparameters such as the number, size, and aspect ratio of the prior box in the one-stage method affect the training and detection speed of the model

[0008] In addition, the backbone network in the existing face detection model is migrated from the general target detection backbone network. The complexity of the model itself is high and the appearance modeling of the face generally contains redundant information. The detection process requires computing power. Supported by powerful GPU, etc., it is still difficult to achieve real-time on general mobile terminals and CPUs

[0009] The slow detection speed of the existing face detection method based on deep neural network on the CPU is mainly caused by the high complexity of the model itself; in the one-stage detection method, the unreasonable setting of the prior frame will also affect the detection accuracy and speed

Images