| Face detection is a basic and important research topic in computer vision which is widely used in intelligent security,human-computer interaction and many other fields.In recent years,with the social and personal security issues become increasingly prominent,intelligent monitoring technology based on face information has obtained widespread concern and rapid development.Face Detection,as a key module of this kind of intelligent video surveillance product,is the basis for the system to realize the recognition of face and early warning function.Therefore,the research of face detection technology in monitoring scene has a strong practical application value.This paper mainly studies the face detection method of station entrance and exit.In order to improve the performance of face detection in this particular situation,in the aspect of model training,a method for adaptively re-training model is proposed;in the aspect of feature representation,a characterization method fusing traditional features and deep features is proposed.The experimental results on the face data set of the station ticket barriers show that these methods improve the accuracy of face detection.Specific research work and innovation are as follows:1)To solve the problem that detectors trained on public data sets only obtain modest performance in the scene of ticket barriers,we collected a lot of data,annotated the training set and test set automatically and manually,and built a face data set in the scene of ticket barriers.The aggregated channel features model trained on the training set of the scene is then taken as the benchmark model of the research work.2)In order to improve the detection accuracy of the aggregated channel features model,we propose an adaptive model re-training method.This method first uses the pre-trained model to detect the face,and then collects the face/non-face samples from the test image without supervision,then updates the training set with these samples,and finally retrains the classifier on the training set.The experimental results show that the accuracy of the re-training model is significantly higher than that of the benchmark model,and thus verify the validity of the training method.3)In order to further improve the face detection results in the scene of ticket barriers,we propose a face representation method which combines the aggregated channel features and deep features.In this method,the candidate bounding boxes are firstly selected by the aggregated channel features model.Secondly,the candidate boxes are classified based on the aggregated channel features and the deep features respectively.Then,the classification confidence scores of different features are fused.Finally,the candidate box is judged according to the threshold.The experimental results show that the fusion of deep feature can effectively improve the accuracy of ACF model. |
