Research On Face Detecetion Based On R-FCN In Complex Scenes

In recent years,face detection technology has become one of the most popular research topics.With the maturity of face detection technology,the scenes used in society can be seen everywhere,such as banks,major shopping malls,customs,road safety,station security checking.However,most of the face detection methods proposed at present are mainly for face images under strong constraints.These face images are mainly based on simple backgrounds and similar faces.However,affected by complex factors such as scale change,atypical pose,and blur,the above-mentioned face detection method has the problems of low efficiency and high missed detection rate in actual scenes.Therefore,this paper studies the face detection method based on regional full convolutional neural network(R-FCN)to improve the face detection effect in complex scenes,that is,in complex scenarios,it not loses important features,but also improves the target of detection accuracy.The main work of this paper is as follows:(1)For the backbone network ResNet-101 of R-FCN,when the input image is convoluted and the information features are extracted,the problem that the important features of the target part are lost may occur.In this regard,this article introduced the Squeeze-and-Excitation module.First,the channel correlation between convolutional layer features is modeled to improve the model's expressive ability;then,the global information is used to selectively enhance the informative features,and this article will be in the final stage of ResNet-101.The expansion convolution is used instead of ordinary convolution to systematically aggregate the context information of multi-scale features.Finally,through experiments on FDDB and WIDER FACE dataset,it is proved that ResNet-101 and Squeeze-and-Excitation modules are merged in this paper.It has a good promotion effect on the performance improvement of the R-FCN network model.(2)In order to detect the face more accurately,this paper sets a smaller size anchor in the R-FCN network,and performs multi-scale inspection inside the network to combat the difficulty of detecting small targets.By reading a large number of documents,the traditional face detection method generally extracts the characteristics of the whole image through a single network structure.When the face changes due to factors such as scale,expression,partial occlusion,etc.,the method shows certain limitations.The reason is that each image sample contains content,but its content is not evenly distributed on every area in the image.To this end,this paper introduces a residual attention mechanism in the R-FCN network structure,and learns as much as possible about each area of the picture.When the image has complex scenes and large appearance changes,the method proposed in this paper can ignore the irrelevant information and focus on the key information,so that the features are not affected by these factors.Finally,the effectiveness and efficiency of the proposed method are verified by experiments on the WIDER FACE dataset.