| With the flourishing development of the domestic transportation industry,traffic congestion and traffic accidents are also increasing.In this context,the intelligent transportation system(ITS)designed to improve transportation efficiency and safety levels has emerged.In the intelligent transportation system,image acquisition devices such as cameras are the main means of obtaining traffic information.However,due to various factors,such as changes in lighting conditions,adverse weather,and other factors,the captured images may suffer from quality degradation issues such as blurring,reduced contrast,and insufficient brightness.With the help of image quality evaluation technology,it is possible to effectively control the image quality of images processed by traffic image enhancement algorithms,providing a better foundation for subsequent processing and recognition.In addition,image quality evaluation methods can not only objectively evaluate the performance of algorithm processing results,but also help guide the optimization of traffic image enhancement algorithms to achieve the best perception level.The main work of this paper is as follows:(1)To address the issue of distorted image quality assessment,a no-reference distorted image quality assessment method based on feature learning in contourlet domain is proposed in this article.The components of CIELAB color space and gradient magnitude image of distorted image are firstly obtained.Next,these feature maps are decomposed by a 3-scale and 8-directional contourlet transform and three types of distortion-aware features are extracted from them: the energy feature of the contourlet subband,the energy difference feature of subbands between scales,and the joint statistical distribution feature of subbands across scales.Finally,the ultimate quality prediction model is obtained by learning the above perceptual features with a support vector regression model.On the synthetic distortion databases TID2013 and CSIQ,the proposed no-reference method Contourlet QA achieves improvement of 7.2% and 5.6% of SROCC results over the second-best method.In terms of average value,Contourlet QA method achieves improvement of 4.0% of SROCC result over the second-best method.(2)To address the issue of dehazed image quality assessment in intelligent transportation,a no-reference image quality assessment method based on feature learning in complex contourlet domain is proposed in this article.Firstly,in order to measure the visibility of the dehazed image,the complex contourlet transform is utilized to extract sharpness-,brightness-and contrast-aware features from dehazed images respectively.Among,the logarithmic energy of the complex contourlet subband is used to calculate the intensity of sharpness,the subband entropy of brightness variation map corresponding to dehazed image is utilized to estimate the intensity of brightness,and the intensity of contrast is calculated based on the statistical difference of the sub-bands between dehazed image and the corresponding contrast variation map.Secondly,to detect the image distortions introduced by the dehazing algorithms,this structure and color-aware features are extracted from dehazed images,respectively.Among,the retention of structural information in the dehazed image is estimated from the LBP histogram statistics of the image after complex contourlet reconstruction,and the naturalness corresponding to the color information of image is measured with the across-scale and across-orientation statistics of the saturation and chroma maps.Finally,the visibility-and distortion-aware features are combined to predict the ultimate quality score of the dehazed image.On the synthetic dehazed image quality assessment databases SHRQR and D-HAZY,the proposed method VDA-DQA achieves improvement of7.2% and 3.0% of SROCC results over the second-best method.On the authentic dehazed image quality assessment databases IVCDehazing and ex Be DDE,the VDA-DQA method achieves improvement of 13.3% and 0.4% of SROCC results over the second-best method.(3)To address the issue of evaluating the quality of image enhancement algorithms such as dehazing,deraining,and denoising in traffic scenes,a no-reference image quality assessment method based on Two-Stream Complex-Valued Convolutional Neural Network is proposed in this article.Considering that visual visibility improvement and scene content preservation are crucial for the quality of enhanced image,the interference perception map corresponding to the distorted image is firstly calculated and then fed into two sub-networks with the distorted image respectively.In each sub-network,a dual-tree complex wavelet transform is used to obtain the corresponding complex-valued response of the real-valued input image.Subsequently,several cascaded complex-valued multi-scale aggregation modules are utilized to extract layer-level perceptual features from the high-and low-frequency components of the complex-valued responses and the global and local perceptual information under each layer are merged with the help of high-and low-frequency fusion module.Then,the high-and low-frequency fused features under different layers are again combined for subsequent processing.Finally,the deep features learned by two sub-networks are further fused and the ultimate quality score is obtained by regression learning with a quality regression module consisting of three fully connected layers.On the dehazed image quality assessment databases DHQ,ex Be DDE and DEHID,the proposed method TS-CVCNN achieves improvement of 1.5%,1.6% and 2.2% of SROCC results over the second-best method.On the deraining image quality assessment database IVIPC,TS-CVCNN achieves improvement of 5.3% of SROCC result over the second-best method.On the denoising part of Ca HDC database,TS-CVCNN achieves improvement of 0.4%of SROCC result over the second-best method.(4)To address the poor generalization performance of existing deep real-valued networkbased image enhancement methods in real-world scenes,a traffic image enhancement method guided by image quality assessment model is proposed in this article.In real traffic scenarios,this method can effectively improve the clarity and recognizability of images,thus providing more reliable visual information for applications such as traffic sign recognition and vehicle reidentification.Firstly,an image enhancement network based on a complex-valued U-Net structure is designed,and a complex-valued transform is performed on the real-valued image before feeding into the network.Among,the encoder part of the enhancement network consists of a complex-valued convolutional layer,several complex-valued residual dense blocks with complex-valued attention,and complex-valued downsampling modules.In the subsequent feature recovery module,several complex-valued residual dense blocks are used to further process the deep features extracted from the encoder.The decoder part uses several complexvalued residual dense blocks,adaptive complex-valued skip connection structure and complexvalued upsampling modules to recover the detailed information of image.In the training stage,this paper introduces the enhanced image quality assessment network in work(3)to calculate the distance difference between the enhanced image and the clear image as the quality assessment loss and jointly optimize the complex-valued enhancement network with the combination of L1 loss and quality assessment loss.On the outdoor part of SOTS,synthetic part of HSTS and the Haze-4K database,the proposed method QGENet achieves improvement of 0.37 d B,0.93 d B and 1.5d B of PSNR results over the second-best method.On the RTTS and the authentic part of HSTS database,QGENet achieves improvement of 15% and 21.9% of noreference quality assessment results over the second-best method.In addition,for real traffic tasks in foggy weather,the proposed QGENet can not only effectively enhance traffic signs,but also improve the accuracy of vehicle re-identification in foggy weather.The above results indicate that the proposed QGENet can ensure the optimization of traffic flow,and reduce congestion and accident rates in intelligent transportation systems. |
PDF Full Text Request