Research On Deep Image Dehazing And Its Evaluation And Passive Detection

Haze is a common weather phenomenon,and the acquired hazy images suffer from effects such as low contrast,color distortion,and blurred details.When hazy images are applied to various computer vision tasks,they will directly affect the performance of image content analysis and understanding such as target recognition.Image dehazing,i.e.,maximally recovering the original image from degraded images with haze,is of great significance for improving the performance of computer vision tasks.In recent years,deep learning-based image dehazing research has received extensive attention,but the proposed deep image dehazing models usually require a large number of image pairs consisting of hazy images and their corresponding clear reference images for training.However,collecting such naturally acquired image pairs is very difficult.For this reason,current deep dehazing models are usually trained using synthetic hazy images,but the existing synthetic hazy images suffer from a lack of depth perception and insufficient diversity,which limit the performance of image dehazing models.Second,image dehazing needs to be evaluated objectively to promote image dehazing research,especially to select suitable image dehazing methods/models according to different applications.In addition,image dehazing is an image processing operation,which can also be considered as an image tampering operation from the perspective of image content security.Therefore,this paper focuses on image dehazing and investigates hazy image generation,natural hazy image dataset construction,deep image dehazing,objective evaluation of image dehazing,and passive detection of image dehazing.Specifically,the main work and innovations are summarized as follows.Firstly,considering that natural hazy images visually correlate fog density with scene depth,a generative adversarial network(GAN)-based hazy image generation model,HazeGAN,is proposed for generating realistic hazy images with scene depth perception.HazeGAN contains a depth estimation network and incorporates an atmospheric scattering model to obtain multi-scale image features,and then embeds these features into a generative antagonistic network to generate depth-aware hazy images.Experimental results show that the hazy images generated by the Haze GAN model have more realistic visual effects as the hazy density changes with the scene depth.Using the synthetic hazy image dataset generated by Haze GAN can provide better dataset support for the training of deep dehazing models,and help improve the robustness and other performance of the deep dehazing model.Secondly,a new naturally collected hazy image dataset,called RW-HAZE,is constructed,including 210 pairs of hazy images and clear reference images.The RW-HAZE is obtained through image acquisition by a camera installed on the top of the mountain for weather monitoring,which can keep the image pairs aligned in physical space,and a clear reference image corresponds to multiple hazy images with different fog concentrations.In addition,by carefully selecting the scene and time interval of image capture,it is guaranteed that the image pair only has background differences caused by slight changes in the weather,and the rest remains as unchanged as possible.The RW-HAZE is significantly different from existing datasets such as Be DDE.It can be used not only to train a deep image dehazing model,and improve its generalization performance on natural hazy images,but also to evaluate the dehazing quality of full-reference images.Third,to alleviate the time-consuming and laborious collection of large-scale datasets for deep image dehazing,a self-supervised learning-based image dehazing framework is proposed,and design a pseudo-label generator to automatically generate training data without additional manual labeling,which greatly simplifies the training of the model.Based on the observation that the distribution of haze in the same scene is variable in a real-world haze environment,we propose to use custom depth maps for synthesizing hazy images.For any input clear image,multiple different depth maps are defined to generate multiple hazy images with different haze distributions to enrich the diversity of training data.In addition,the proposed self-supervised dehazing framework is general,and most deep learning-based dehazing models are able to be trained self-supervised under it without manually collecting training data.Experimental results show that the dehazing performance of SSID far exceeds existing unsupervised dehazing models and is comparable to fully supervised dehazing models.Fourth,combined with the domain knowledge of image dehazing,a full-reference quality evaluation model Deh IQA for dehazed images based on transfer learning is proposed.The source task of transfer learning is to use the classification network to classify the clear image and the hazy image,so that the classification network learns the features haze-related through training.The objective task is to extract the features of the hazy image and the reference image using the feature extraction layer of the classification network and perform a similarity measurement to obtain a quality factor of the dehazed image.In addition,considering that the haze distribution of real-world hazy images is not uniform,the larger the haze concentration of the original hazy image is in the region,the more likely the dehazing process is to produce distortion.For this reason,a haze attention module is proposed so that the model focuses on the regions in the dehazed image where distortion may exist.The model can solve the problem that the deep dehazing image quality evaluation model requires a large amount of label data.Experimental results show that the evaluation results of Deh IQA are more consistent with the quality of human visual perception and more robust than the existing evaluation metrics such as Realness Index and Visibility Index.Fifth,image dehazing operation is usually used to enhance the visual perception quality of the image,but it can also be used to maliciously tamper with the image,which subverts people’s concept of ”seeing is believing”.By analyzing the formation mechanism of hazy images and the characteristics of dehazed images,it is observed that the illumination consistency of images is destroyed during the dehazing process and artifacts are left behind,which are magnified in the inverse-intensity chromaticity space.To this end,a passive detection dual-stream model is proposed,called DDNet,for image dehazing operation,which performs feature extraction by RGB stream and inverse-intensity chromaticity stream separately.In addition,to improve the complementarity of the dual-stream features,an adaptive feature fusion method is designed,which can adaptively adjust the weight coefficients of the dual-stream network using the loss values of the single-stream network as feedback.Experimental results show that DDNet not only achieves high detection accuracy of image dehazing operation,but also has a strong robust performance.