Research Of Facial Feature Detection Based On Random Forest

Facial feature detection is an important issue in Computer Vision. Based on face recognition, face tracking as well as facial expression recognition, it can be widely used in various areas, for example, medical diagnosis, automatic identity identification, video tracking, human-computer interaction and military analysis. The main task of facial feature detection is to locate facial features quickly and exactly in images for further analysis, including the eyebrows, the eyes, the nose and the mouth. Recently, coincident with the rapid development of IOT(Internet of Things) technology, as well as the increasingly deeper research conducted by scholars from all over the world, great improvements have been made in facial feature detection, a lot of performances having been yielded at the same time. However, due to the complexity of face images as well as the variance of head poses, several aspects of the issue remain to be improved, including the real-time performance, accuracy and robustness.Centered on the detecting methods based on random forests, the thesis proposes several improved methods, in order to deal with various head poses as well as facial occlusion of ATM terminal. The thesis mainly consists of three parts:Firstly, the facial feature detection based on random forests is analyzed in detail. In the stage of training, the model of random forest can be constructed with the facial image database. Then, the image databases including LFW, LFPW and COPW are used to test. Finally, the problems and flaws of the proposed method are detailedly analyzed according to the tested results.Secondly, In order to solve the problem of various head poses for facial feature points detection, the Revised Structured-Output Regression Forests(RSO-RF) is proposed in this paper. In the training stage, different head poses are categorized into five sets, including the left profile, left, front, right, and right profile. And all the categorized head poses are employed to construct different models, respectively. Meanwhile, the estimated head pose model is also constructed based on the random forests. In the testing stage, the head pose is primarily estimated by the estimated head pose model, and the random forest detection model then can be selected from the corresponding head pose for the facial feature points detection.Thirdly, to solve the problem that criminals are occluded in the existing ATM terminal, this paper proposed a partial occluded method based on RSO-RF for facial feature detection, which is integrated with the occlusion information in RSO-RF. In the stage of training, information of occlusion is marked for all training data, which fall into three categories, 'No occlusion', 'Occlusion on eyes' or 'Occlusion on the mouth'. When constructing a model, information of both occlusion and features will be seen as the votes and stored at leaf nodes. In the testing stage, the image is whether occluded can be estimated based on the occlusion estimation method based on random forests. For the partially occluded images, the no-occluded facial feature points are located.Finally, experiments on LFW and LFAOL demonstrate that the proposed various head pose facial feature detection method based on SO-RF can obviously improve the accuracy of detection and decrease the error rate for various facial images; Moreover, experiments on AR and LOFAO also demonstrate the proposed partially occluded facial feature detection method based on RSO-RF can not only detect whether the testing images are occluded, but also locate the no-occluded facial features well.