Research On A Ranking-driven Method For Evaluation And Improvement Of Underwater Image Quality

In recent years,deep learning technologies have dominated the field of image processing.However,it is difficult to collect paired images in an underwater environment,which greatly limits the application of deep learning in the field of underwater image quality evaluation and quality improvement.To address above problems,in the framework of deep learning,this dissertation proposes a non-reference underwater image quality evaluation network that combines the idea of ranking.To handle both image quality evaluation and improvement,the proposed evaluator is further used to improve the underwater image quality.Therefore,the main work and innovations of this dissertation are as follows:First,a new non-reference underwater image quality evaluation network is proposed.By subjectively scoring the images in existing underwater datasets,the problem of underwater image quality evaluation is transformed into a ranking problem,and the corresponding neural network structure is designed as a ranker.At the same time,a differentiable ranking loss function is designed based on the subjective score and used to train the ranker.In this way,the trained ranker can effectively evaluate the underwater image quality consistent with subjective perception;Second,a new underwater image quality improvement network is proposed.In order to effectively deal with the degradation of underwater image quality,based on the idea of divide and conquer,this dissertation separately processes the color average vector and image details.In this way,the learning problem is greatly simplified.At the same time,a new adjacently connected network structure is proposed,which can effectively explore rich feature representations for underwater image quality improvement.In addition,by using the above-mentioned ranker as a loss function,the quality improvement network can be guided at the semantic level during the training process,so that the improved underwater image quality is more in line with subjective perception.Experiment on both synthetic and real-world underwater images shows that the quality evaluation network proposed in this thesis can be directly used as an evaluation index in the field of underwater image processing.In addition,the proposed underwater image quality improvement network can achieve effective color correction and contrast enhancement on both synthetic and real-world underwater images,which can significantly improve subjective visual effects and objective indicators.