A Study On Person Re-identification,Image Reflection Removal And Low Light Image Enhancement Using Convolutional Neural Networks

In recent years,with the rapid development of society,people are about to usher in a fast-developing intelligent society.Computer vision tasks such as target detection,target recog-nition play an important role in video surveillance.Person re-identification,image reflection removal,and low-light image enhancement as sub-tasks of computer vision area have at-tracted much attention by researchers and have become a hot issue.Person re-identification refers to the cross-cameras pedestrian retrieval task,which identifies a specific person from the same scene or different scenes.This task shows superior potential for large-scale video surveillance.Among the research goals of pedestrian recognition,there are a variety of challenging issues,including low resolution of video surveillance,illumination variation,occlusions,and viewpoint changes.Since computer vision tasks are mostly based on images and the quality of images collected in different environments are different,image degrada-tion often seriously affects the performance of subsequent subtasks.When we need to take pictures through the glass,the collected images often suffer from the reflection,which leads to the deterioration of the image quality.In reality,the reflection and background images are often natural images,and they all follow the same distribution.Therefore,it is an ill-posed problem to obtain a clean image(i.e.,a background image)from a single reflection map.In reality,sometimes we also inevitably encounter low-light environments.At this time,the captured images suffer from noise and color distortion.Low light enhancement refers to enhancing such images to produce a noise-free image.Therefore,this thesis conducts in-depth research on person re-identification,image reflection removal and low-light image enhancement,and mainly discusses how to apply deep learning techniques to these fields.The main work of this paper is listed as follows:1.An unsupervised person re-identification algorithm is proposed.Most of the current widespread person re-identification algorithms are based on a supervised framework.How-ever,supervision means that a large amount of labeled data is required,which needs a lot of manpower and material resources.Although the current unsupervised algorithm is still poor compared with the supervised algorithm in performance,its advantage of being scenery unlimited and not limited to data is obvious.Therefore,we propose an unsupervised person re-identification algorithm based on soft labels and reliable labels,which can make full use of all data information to obtain robust features.2.A multi-modal reflection removal algorithm is proposed.The reflection removal study of a single image is often difficult to achieve the desired effect,mainly because the back-ground image and the reflection image are consistent with the natural image distribution,and the known information is too small.In a series of observations,we found that the depth image obtained by Kinect is robust to reflection at a certain angle,that is,the depth image can be considered to be unaffected by reflection within a certain angle range.Based on this,we designed a multi-modal reflection removal framework based on the depth image and re-flection image,and used the depth map as the guiding information to do reflection removal.Except training and testing on the public data,we also take a total of 205 sets of reflected images in a real scene to evaluate the performance under real scenes.3.A low-light image enhancement algorithm based on image gradient is proposed.Since low-light images often contain a lot of noise,some details of the image are often lost during image recovery.We designed a concurrent network framework to learn the image gradient while restoring the enhanced image.The gradient information of different scales is trans-ferred to the image recovery network,thereby obtaining a detailed picture with less noise.To successfully restore color well,we further build a C-CNN for color restoration.