Research On Unsupervised Person Re-identification Algorithm Based On Deep Learning

Person Re-Identification(Re-ID)technology has always been a key task in the field of computer vision and pattern recognition.This technology is generally regarded as a subset of image retrieval.Person re-identification is a specific recognition technology that aims to locate a specific pedestrian in a cross-camera video or image given a unique pedestrian.This technology is usually combined with face recognition,pedestrian tracking and other technologies,combined with applications in the field of video surveillance and smart security.In the cross-domain person re-identification methods based on pseudo-label predicting,the performance of the model largely depends on the quality of the pseudolabel,and the outliers generated by clustering contain a wealth of knowledge.An urgent problem is how to improve the quality of pseudo-labels in iterative training and how to leverage abundant outliers.Secondly,mining more accurate positive samples to improve the learning ability of the model has always been a common practice in the person re-identification algorithm.How to make the model mine more accurate positive samples,and design the model framework and loss function to improve the model's ability to learn discriminative features It is another problem.In response to the abovementioned first problem,this paper first proposes a cross-domain person reidentification algorithm based on multi-clustering.This algorithm uses a variety of clustering algorithms to assign pseudo-labels,generates multiple pseudo-labels for each sample by clustering algorithm,and reassigned pseudo-labels for outliers,combined with a multi-branch network,gradually improves the quality of pseudolabels and the recognition performance of the model in the iterative process.In response to the above-mentioned second problem,based on the multi-cluster cross-domain person re-identification framework,this paper proposes a cross-domain person reidentification algorithm combining hard sample mining and multi-clustering.The algorithm is based on data augmentation technology,through the combination of pseudo-label and cross-selection to mine hard samples,and gradually improve the learning ability of the model in the iterative process.In summary,the main contributions of this paper are as follows:(1)This paper proposes a cross-domain person re-identification algorithm based on multi-clustering to solve the problem of poor label quality in the iterative training process.First,use the labeled source domain dataset to training the basic model,and use the basic model to extract the global and local features of the unlabeled samples;then,calculate the distance between feature pairs,use different clustering algorithms for different features,and assign pseudo labels;secondly,assign outliers to the most similar clusters;finally,use pseudo-labels as supervised information for supervised learning,so as to realize the cross-domain recognition ability of the model.The experimental results show that the cross-domain person re-identification algorithm based on multi-clustering proposed in this paper can achieve better recognition results.(2)This paper proposes a cross-domain person re-identification algorithm that combines hard sample mining and multi-clustering to solve the problem of mining hard samples to improve the model's ability to learning discriminative features.First,based on the multi-cluster cross-domain person re-identification framework,the network is divided into original branches and enhanced branches;then the difficult samples are selected by cross-selection;finally,the model is trained by combining pseudo-labels and selected hard samples until the depth model is stable.The experimental results show that the cross-domain person re-identification algorithm proposed in this paper,which combines hard sample mining and multi-clustering,can achieve a significant improvement in evaluation indicators.