Research On Face Accessories Detection Based On Faster-RCNN

Face attributes recognition and detection is a high-profile research topic in the field of computer vision,and lots of scholars have done a full research on this.Face attributes include internal and external attributes such as ethnicity,gender,posture,age,accessories,expressions,etc.In some large scenes(shopping malls,libraries,etc.),The staff needs to backtrack video materials to lock in some suspicious characters.Face accessories play an extremely important role in assisting the staff backtracking the target.This thesis uses deep learning-based Object Detection to optimize the detection of face accessories.At present,the Object Detection network framework based on convolutional neural network can be divided into two categories: one-stage method and two-stage method.This thesis proposes two improved algorithms based on the typical algorithm Faster-RCNN in the two-stage method.(1)Faster-RCNN selects the top N region proposals with the best scoring effect through non-maximum suppression for subsequent training.During training,the Regional Number Adjustment Layer is introduced to judge the current training effect in real time,and the number of N is appropriately increased or decreased according to the current training effect.N converges with the convergence of the overall network.The number of region proposals obtained through training can better balance between the detection speed and average accuracy of the model.In addition,deep convolutional neural networks are prone to degradation in training,and the introduction of residual network can effectively suppress this phenomenon.Based on Resnet50 for improvement,a 58-layer feature extraction network was rebuilt.The experiment was conducted on the PASCAL VOC 2007 data set.Compared with the classic network model,the speed rate was increased by 18%,and the detection rate was increased by 3%.(2)In addition,make improvements to the detection of face accessories under the requirements of specific scenes,such as Multi-scale Training and rich anchor frame styles.On this basis,a face accessory detection method based on small target is proposed.Since FasterRCNN only uses the last feature map of the low-level feature extraction module,the detection rate for small target is not high.The feature extraction network is divided into five convolution modules,and the feature map obtained by each convolution module is input into the Regional Proposal Network in parallel to generate region proposals,making full use of the features of different levels of the convolution network,and improving the overall network's ability to small target Robustness and detection rate.