Research On No-reference Underwater Image Quality Evaluation And Its Benchmark Datase

The ocean contains huge resources and is regarded as the sixth continent that can be used by human beings.It is a far-reaching task to develop and protect the marine resources.Underwater images are widely used in the exploration and development of marine resources because of their advantages of easy acquisition,low cost,and high information density.However,due to the selective absorption and scattering effects of water medium,the captured underwater images usually have color distortion,low contrast,blur and other distortions,which seriously affect the interpretation of image content.Therefore,how to evaluate the quality of underwater images more accurately is of great significance for the design of underwater image enhancement and restoration algorithms and high-level vision applications.To solve the above problems,a no-reference underwater image quality evaluation method based on quality-aware features is proposed,called NCCSS.NCCSS extracts a set of 56-dimensional feature vectors from five aspects: naturalness,color,contrast,sharpness,and structure.Specifically,considering the color degradation problem of underwater images,the color-cast weight colorfulness measurement and color consistency are proposed to capture color information.In addition,considering the visual saliency,a saliencyweighted contrast measurement is designed to evaluate the contrast information more accurately.After that,NCCSS extracts all the aforementioned quality-aware features from a training set to build a prediction model which maps these features onto the image quality level by using SVM(Support Vector Machine),and then employ it to predict the quality of an image.Experimental results show that,compared with the commonly used methods,NCCSS achieves high consistency with human subjective perception.The lack of publicly available datasets containing underwater images is a hindrance to the development of the field of underwater image quality evaluation.In order to address this issue,a large-scale underwater image dataset called UID2021 is established.This dataset comprises of 60 multi-degraded underwater images from various sources,covering six common types of underwater image degradation.Additionally,900 quality improved versions of the 60 multi-degraded images are generated by using 15 of the most commonly used underwater image enhancement and restoration algorithms.MOS(Mean Opinion Score)ratings were obtained from 52 observers with diverse backgrounds using a pairwise comparison sort method.The UID2021 dataset serves as a valuable addition to the existing underwater image datasets for image quality assessment,and it can be used as a new benchmark for evaluating the performance of no-reference underwater image quality assessment algorithms.