| In recent years,unsupervised pedestrian detection has been developed.In order to improve the detection performance,a method using an unsupervised image translation framework to generate an intermediate domain dataset between the source and the target domains is proposed.The images in the intermediate domain dataset generated by this framework not only retain the content characteristics of the source domain,but also have the style characteristics of the target domain.Then replace the source domain dataset with the generated intermediate domain dataset to perform cross-domain pedestrian detection in the target domain.Since the differences in scene distribution between the intermediate domain and the target domain are relatively smaller,the performance of cross-domain detection is usually improved.However,although the intermediate domain obtained by using the unsupervised image translation framework can improve the performance of unsupervised pedestrian detection,due to the instability of the internal generation network of the unsupervised image translation framework,some images have varying degrees of distortion in the newly generated intermediate domain dataset.This undoubtedly has an adverse effect on the performance of cross-domain detection.Therefore,this paper conducts research on issues such as improving the quality of intermediate domain datasets generated by unsupervised image translation techniques with the aim of improving the performance of unsupervised pedestrian detection.The paper work consists of two main aspects.To address the problem of efficiently acquiring high-quality intermediate domain datasets,this paper proposes an end-to-end intelligent high-quality image translation method.Chapter 3 of this paper shows that the method not only obtains extremely highquality intermediate domain datasets,but also effectively improves the performance of unsupervised pedestrian detection.To address the problem of obtaining high-quality intermediate domain datasets without reference images,this paper proposes a nonreference high-quality image translation method based on high-quality filtering.Chapter4 of this paper shows that this method can directly acquire high-quality intermediate domain datasets without reference images and effectively improves the performance of unsupervised pedestrian detection.This paper first proposes an end-to-end high-quality-awareness image translation method.The main method is to select a suitable reference image quality assessment index and make appropriate adjustments so that it can be integrated into the total loss function of the unsupervised image translation framework as a new loss.At the same time,rationally adjust the constraint factor of the new loss item to control the degree of influence in the image generation process,thereby controlling the quality of the generated images,and directly generating a high-quality intermediate domain dataset.This method has a constraining effect on the image quality during the generation of the intermediate domain images,not only can obtain extremely high-quality intermediate domain datasets,but also can significantly improve the training speed of the unsupervised image translation framework.Based on the experimental results,it can be concluded that the method is not only efficient in obtaining high-quality intermediate domain datasets,but also effective in improving the performance of unsupervised pedestrian detection.The first method in this paper achieves good results and further hopes to obtain the same high quality intermediate domain datasets without reference images and improve the performance of unsupervised pedestrian detection effectively.So this paper then proposes a high quality image translation method based on high quality filtering without reference.The main objective is to directly control the image quality without reference images.The main method is to evaluate and filter the datasets twice by selecting appropriate non-reference image quality assessment indexes.The first time is to evaluate the original datasets.Retain the high-quality part and remove the low-quality part of the images,which are used for the training of the unsupervised image translation framework.The second time is to evaluate the generated intermediate domain datasets.Retain the high-quality part,and replace the low-quality part of the images with the corresponding original images,so as to prepare the intermediate domain datasets and use it for unsupervised pedestrian detection.Based on the experimental results,it can be concluded that the method effectively improves the performance of unsupervised pedestrian detection.For example,it can be achieved to reduce the original unsupervised pedestrian detection miss rate of 19.13% and the unsupervised pedestrian detection miss rate of13.87% using the unsupervised image translation framework to the unsupervised pedestrian detection miss rate of 10.69% for the proposed method in this paper under the miss rate metric. |
