Research On Person Re-identification Algorithm In Low Quality Conditions

Person re-identification is a technology to search a pedestrian under the condition of cross-camera.Compared with face recognition,it is more practical.The research scenarios and datasets are closer to reality,so low-quality issues is considered inevitably.For example,label in low-quality(label noise)caused by mislabeling and low image quality(image noise)caused by poor video acquisition conditions.In the past two years,as person re-identification is researched in the scene of large-scale real-world application,these two problems are researchers really need to face.As for the label noise,few researcher have discussed it in this research area.However,the research method is slightly different in the context of image classification tasks.Because,the gap of image quality is obvious,and the quality in image classification is better;there is a large gap in the amount of data.The dataset of the person re-identification is generally in the unit of 10,000,while the other is generally in the unit of tens of millions.For the problem of image noise,only single noise type was considered before,and the noise types in real-world application scenarios are generally mixed.To this end,this dissertation proposes methods for these two issues:(1)Learning with Label Noise Based on Self-adaptive Person Re-identificationIn the past two years,many person re-identification methods have achieved high accuracy on benchmark,and their goals include learning sufficient feature representations based on given samples.These methods are based on a premise: the labels of image are accurate.However,label noise can be found on academic datasets,it is also present in large-scale datasets inevitably when person re-identification enters large-scale real-world application scenarios.To this end,this dissertation proposes a self-adaptive person reidentification method with label noise.In the early stage of training,when the non-noise samples(the sample label is correct)are fully learned and the noise samples(the sample label is wrong)are insufficiently fitted,the non-noise samples are filtered according to the distribution of training loss.There are some hard samples within the non-noise samples,and they have the characteristics of difficulty in fitting,same as noise samples.For this reason,weights of them are resorted.As the iterative process progresses,the hard samples get relatively larger weights while the weights of noise samples gradually attenuate.In the overall model iteration process,the filter learning module distinguishes between noise and non-noise samples,and the revise learning module minimizes the damage suffered from the noise samples,and makes up for the problem of insufficient learning from hard samples.(2)Learning with Image Noise Based on Self-adaptive Person Re-identificationPerson re-identification is a research field that is close to reality.There are widespread problems of poor image quality caused by low resolution,missing parts,blur and other factors in the dataset.Previously,separate methods for low-resolution problems have been proposed,but the image noise types in the dataset are complex.To this end,this dissertation proposes a person re-identification method that faces mixed noise types and improves the ability of feature of low-quality image.Specifically,as normalizing the feature representation to the unit hypersphere of the corresponding feature space,it alleviates the problem of more similarity between low-quality images and inter-classes.In addition,in order to improve the correlation between the feature representation and the features in the foreground information,a position attention mechanism is exploited in the feature representation extraction process.