Visual Computation Model And Methodology For Degraded Image

Image is an important source of information for the human eyes and computer vi-sion system to perceive the outside world and obtain information from the outside world.The quality of the image directly determines whether the image can be recognized by the human eyes and computer vision system.However,the imaging process will in-evitably be affected by various degradation cues(e.g.,object motion,camera shaking,poor lighting condition,bad weather,etc.),resulting in image quality degradation.On the one hand,image degradation will cause image distortion and partial loss of infor-mation,reducing the visual perception effect of the human eyes;on the other hand,the degradation process will destroy the statistical properties and structural information of image,which will seriously affect the visual recognition of image,leading to computer vision system performance decrease.To enhance the quality of degraded images and improve human visual perception,researchers have conducted extensive research and proposed numerous image enhance-ment algorithms.However,most of the existing image enhancement algorithms are devised to address the single-type degraded image enhancement problem.In the practi-cal environment,the captured degraded images often contain multiple degradation cues at the same time,and there may be mutual influences among the multiple degradation cues.How to design a general image enhancement strategy that can accurately model multiple degradation cues in the image,and how to suppress the coupling relationship between multiple degradation cues during the modeling process,is still to be resolved urgently research topics.To improve the recognition performance of computer vision system in degraded conditions,researchers have carried out extensive research and proposed many solu-tions,which can be summarized as:collecting labeled samples,first enhancing and then recognizing,cascading image enhancement network and classification network for joint training.However,the above methods have a series of shortcomings,such as large-scale labeled samples are difficult to collect,image enhancement algorithms cannot guaran-tee the structural similar regions can be consistently enhanced,joint training will change the parameters of pre-trained models.Under the above-mentioned research background,this paper conducts in-depth research on the problem of hybrid degraded image en-hancement and degraded image recognition,hoping to play a certain role in promoting the research of the visual computational methods for degraded images.The main research contents and innovations of this paper are as follows:(1)Hybrid degraded image enhancement based on image statistical modeling.In the process of hybrid degraded image enhancement,the gradient changes within local regions of image will affect the statistical feature representation of network and reduce the modeling accuracy.To suppress the influence of gradient information on statistical modeling,this paper proposes a statistical modeling method based on pixel disruption strategy,which can destroy the image gradient by randomly disrupt pixels while keeping the pixel statistical properties unchanged and make it easier for the network to extract the statistical properties of the degraded clues.Finally,we take the underwater image enhancement problem as an example to verify the effectiveness of proposed method.(2)Hybrid degraded image enhancement based on fully 1*1 cyclic decoupling network.In the process of hybrid degraded image enhancement,there may be a cou-pling relationship between multiple degraded cues.Due to the influence of coupling relationship,the modeling process for a specific degradation cue will be inevitably be affected by other degraded cues.To address this problem,this paper proposes a cyclic decoupling-based hybrid degraded image enhancement method,which perceives the degradation properties by extracting the statistical characteristics of the degradation cues,and then decouples the coupling relationship through the iterative optimization mechanism.Benefited to the cyclic decoupling network,the mutual influence between degradation cues is suppressed,and the accuracy of network modeling can be improved.Besides,to improve the computational efficiency of network,we first propose a fully 1*1 convolution-based statistical modeling method,and design two fully 1*1 convolution modules to equivalently replace the K*K convolution operations in existing convolu-tion neural network.Finally,we take the low-light image enhancement problem as an example to verify the effectiveness of proposed method.(3)Deep degradation prior for linear low-quality image classification.Degradation clues will destroy the statistical properties of image pixels and cause the degradation of classification performance.To address this problem,a deep degradation prior based linear degraded image classification method is proposed.We find that the distribu-tions of corresponding structure-similar patches in low-and high-quality images have uniform margins under the same degradation condition,and the degraded features will have uniform drifting under the same linear degradation conditions.We propose a fea-ture de-drifting module inspired by non-classical receptive field to learn the mapping between degraded features and clear features,which can be trained without the supervi-sion of semantic labels.Finally,we take the three types of linear degradation(fog,low contrast,and overexposure)as examples to verify the effectiveness of proposed method.(4)Adaptive feature de-drifting based non-linear degraded image classification.The structural and statistical properties of different regions within non-linear degraded images will suffer different degrees of influence,resulting in non-uniform feature drift-ing,which will drop the image classification performance.To address this problem,this paper proposes an adaptive feature de-drifting method for non-linear degraded im-age classification,which exploits the frequency distribution within each region to esti-mate a feature deviation map and leverage it as guidance for feature de-drifting,so as to improve the performance of image classification.Finally,this paper takes the JPEG compressed image classification as an example to verify the effectiveness of proposed method.