Research On Preventing Overfitting Of Person Recognition Network

Person reID(person re-identification)has great significance to public safety supervision.It is still in the laboratory research stage,and the over-fitting problem is one of the most important issues that hinder the development of person reID.At present,the effective solutions are mainly data enhancement and regularization.The former usually uses simple flipping,erasing,multi-scale input and other operations to achieve a limited improvement in network performance.Regularization method such as Dropout,using random zeroing out of some features to enhance the independence between features.However,the random zero operation ignores the difference in the contribution of each feature to the network recognition performance.For the above two questions,this paper proposes new regularization and data enhancement from the perspective of precise intervention network training and data style conversion to improve the generalization of the network more effectively.The specific research content is as follows:(1)Drop Easy-based person reID method is proposed in this paper.In this paper,features are classified into discriminative and indiscriminative ones.Firstly,Drop Easy1 d is proposed in the fully connected layer,according to the distance between the feature vectors of positive or negative sample pairs wherein the discriminative features are zeroed out,while the indiscriminative features are reserved,and the network only learns through indiscriminative features.Furthermore,because networks are always inclined to make up for incomplete information by drawing on the surrounding features in the feature maps,Dropout loses its effectiveness for the network-constraints.To solve this challenge,the Drop Easy2 d method that can be effectively applied to convolutional layers is further proposed in this paper.Drop Easy2 d searches discriminative feature areas in the feature maps by sliding and zeroing windows while reserving the indiscriminative features areas to constrain network learning.The experimental results show that the proposed method can improve the network performance during the extraction and generalization of the discriminative features.(2)A person reID method based on region of interest erasesing(ROIErase)is proposed.Firstly,according to the attention mechanism,the CAM algorithm is used to find the region of interest during the first forward propagation of the network Then use the bilinear interpolation algorithm to extend the feature map to the same size as the input image.Then,according to the position of the region of interest,the input image is erased pixel by pixel,and finally the second forward propagation and reverse gradient propagation processes are performed to update the weight.The experimental results show that the accurate erasing of pictures is more effective than random erasing.The approach potentially drives the network to mine deeper discriminative details that may be overlooked,which improves the intervention effect and intensity of intervention.So it ultimately improves the discriminating power and generalization performance of the network.(3)A data augmentation method based on multi-style transfer is proposed.Firstly,Generative Adversarial Network(GAN)is constructed,and the supplementary dataset is transformed according to various styles of the training dataset,aiming at alleviating the deviation of the feature space distribution between different datasets,thereby achieving the purpose of expanding the dataset.Then a concept of data screening is proposed,and a filter is constructed to filter out the images with low transfer quality,aiming at eliminating redundant transfer pictures and improving the training efficiency of the network.Finally,the person reID network is established,which will use expanded.dataset to train.The experimental results show that the proposed method can effectively capture various styles of dataset and transform them effectively,thus effectively reducing the deviation of feature space distribution between different domains,so as to expand the data and solve the overfit problem of the network.