Person Re-Identification Methodology For Degraded Images

In recent years,the rapid development of intelligent visual perception technology based on deep learning has facilitated the widespread deployment of intelligent surveillance systems in various public places,providing infrastructure support for the construction of safe cities,intelligent transportation,and intelligent security.Person re-identification technology is an indispensable component of intelligent surveillance systems,aiming to identify and match the identities of pedestrians in different camera views within a large-scale surveillance network,thereby enabling fast and accurate retrieval of target individuals.However,due to the influence of imaging environments,the image data obtained by surveillance cameras in all-weather 24/7 scenarios often suffer from various quality degradation phenomena,such as blur,insufficient lighting,occlusion,etc.These phenomena present three challenges to existing person re-identification methods.(1)Degradation factors such as insufficient lighting,low resolution,and adverse weather interfere with the visual appearance of pedestrian images,leading to deviations in the extracted visual features and ranking errors of re-identification results.(2)The image degradation process in the physical world encompasses a wide range of variations and types.Training data collected and annotated manually may not comprehensively cover the distribution of degraded images,thus limiting the generalization ability of person reidentification models.(3)It is difficult to accurately define and quantify the types and degrees of image degradation in real-world scenarios,as well as provide corresponding supervision information through manual annotation.Existing person re-identification methods are only applicable to ideal imaging conditions and struggle to address the aforementioned challenges.When deployed in complex and variable image degradation scenarios,they are prone to retrieval errors,posing significant security risks to intelligent surveillance systems.To address the aforementioned issues,this dissertation explores person reidentification in image degradation scenarios and conducts a series of research efforts:(1)To address the issue of visual feature deviation in low-light conditions,a Retinex-based illumination-invariant person re-identification method is proposed.The method utilizes a pre-trained Retinex illumination decomposition network to enhance the visual quality of low-light images through a bottom-up attention mechanism.It extracts illumination-invariant features from the reflectance maps and combines them with the illumination-enhanced features to form a comprehensive representation of the pedestrian.By achieving cross-illumination alignment at both the image level(illumination enhancement)and the feature level(illumination invariance),this method effectively mitigates the visual feature deviation in low-light degradation scenarios and improves the generalization capability of the re-identification model in the presence of complex illumination variations.(2)To address the issue of visual feature deviation in general degradation scenarios,a degradation-invariant person re-identification method based on disentanglement representation learning is proposed.The method extends the concept of Retinex illumination decoupling to general degradation scenarios,such as low-resolution scenes,and designs a content-degradation decoupling framework based on disentanglement representation learning.By employing adversarial training and feature swapping-reconstruction strategies,the effective decoupling of pedestrian content features and degradation interference features is achieved.Furthermore,by removing the degradation interference components,the method alleviates the problem of feature deviation in general degradation scenarios,significantly enhancing the retrieval accuracy and robustness of the person re-identification model.(3)To address the issue of uneven distribution of high/low-quality images in general degradation scenarios,a training data augmentation method that combines image degradation modeling and adversarial attacks is proposed.The method utilizes a pretrained content-degradation decoupling network to generate degraded images through two algorithms:degradation re-sampling and degradation adversarial attacks.These algorithms generate a large number of challenging samples with consistent identity semantics but varying degrees of degradation.This approach significantly enriches the diversity of training data,alleviating the problem of uneven distribution of high/lowquality images in existing datasets.It avoids the additional cost associated with data annotation and mitigates potential privacy and security risks.(4)To address the issue of missing supervised information in degraded images in real-world scenarios,a degradation invariance learning method based on collaborative supervision training is proposed.The method introduces a unified degradationinvariant representation learning framework that can simultaneously train pixel-level aligned strongly supervised degraded image pairs,image-level aligned weakly supervised degraded image pairs,and unsupervised real image pairs for joint training.This effectively mitigates the problem of missing supervised information in real-world scenes.Furthermore,the use of pseudo-label estimation strategies and image quality ranking loss enables the model to directly extract degradation priors from real images,enhancing its generalization capability in various types of real degradation scenarios.The research in this dissertation effectively improves the ability of pedestrian identity features to resist real-world image degradation perturbations,enhances the retrieval performance of person re-identification models in various degradation scenarios such as weak illumination,low resolution,and adverse weather,and reduces the potential security risks in intelligent video surveillance applications,with broad application prospects and important social value.