Face detection system and method for multi-pose faces

Abstract

The invention provides a face detection system for multi-pose faces. The face detection system comprises a front end detector and a back end detector. The front end detector comprises at least one layer of classifier. Each layer comprises at least two parallel first type classifiers for different poses of faces, wherein the first type classifiers are used for distinguishing candidate faces and non-face windows. The back end detector comprises a second type classifier using a depth nerve network. The second type classifier is used for further distinguishing faces and non-faces in detection results of the front end detector. Accordingly, the invention also provides a face detection method. While the detection precision is improved, calculation expense in the detection process is effectively reduced and detection speed is effectively increased.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to face detection technology. Background technique [0002] The goal of the face detection task is to use a machine to automatically determine whether there is a face in the image or a sequence of images for any given image or a sequence of images, and if there is a face, find its position and size. The problem of face detection is usually abstracted as a binary classification problem, that is, to distinguish between human faces and non-human faces. The classic face detection method mainly learns a classifier based on face and non-face image samples, and then uses the trained data learning classifier to classify each image window through a sliding window on the input image. [0003] Face images will show huge differences in appearance and shape due to internal factors such as age, gender, race, weight, and different head postures of different individuals, while illuminat...

AI Technical Summary

Problems solved by technology

[0004] However, there are still varying degrees of deficiencies in the prior art methods for organizing multiple classifiers

The parallel classifier structure trains a face detector for each pose, and it needs to implement face detectors for all poses for most of the candidate image windows, so it has low detection efficiency, and the false detection of the entire detector The rate increases with the increase of parallel detectors; the pyramid classifier structure can be regarded as a parallel structure sharing high-level nodes in form, except for the first layer, each layer adopts the form of multiple models in parallel , so it has the same problems as the parallel structure in terms of time efficiency and false detection rate, the difference is that the upper layer reduces part of the computational overhead by sharing nodes, and uses a single model to model multi-pose faces There is bound to be a contradiction between the recall rate and the non-face window filtering ability, which will result in a decrease in detection accuracy or an increase in computational overhead, or both; the classifiers on each layer of the tree classifier structure are connected in parallel to The pose is branched from coarse to fine, because its top-level nodes often need more features, and the pose needs to be estimated explicitly or implicitly when branching, and its detection accuracy is heavily dependent on the accuracy of the pose estimate degree, inaccurate or wrong pose estimation can easily cause missed detection and reduce the recall rate of the detector

Some tree detectors have designed a multi-branch mechanism to reduce the dependence on the accuracy of pose estimation, but this does not fundamentally solve the problem, and the multi-branch mechanism itself is a difficulty

[0005] To sum up, for face images with multiple poses, classifiers with low model complexity lack sufficient modeling capabilities, and classifiers with high model complexity have high computational overhead. Currently, multiple classifiers are used. It is also difficult for the method to take into account strong modeling capabilities and small computational overhead, and it is difficult to achieve good performance in both accuracy and speed

Images