Toward a blind quality assessment for underwater images

Underwater images quality assessment (IQA) plays a vital role for some image-based applications. However, existing specific no-reference underwater IQA metrics mainly concentrate on considering visual quality-related features, such as colorfulness, sharpness and contrast, which are insufficient to characterize the image quality comprehensively, resulting in an unsatisfactory prediction performance and limited generalization capability. In this paper, we present a new blind multiple features evaluation metric (MFEM) by extracting five types of features to quantify underwater image quality, including color, sharpness, luminance, structure and texture. Specifically, we combine the colorfulness, color difference and color saturation to represent color features. The sharpness feature is generated by incorporating a saliency map into the sharpness measure. Moreover, based on the natural scene statistics (NSS) regularity, we utilize the NSS features acquired from the illuminance map to characterize the luminance change of distorted underwater image. In addition, the structure and texture features are calculated by employing the gradient-based local binary pattern operator and gray-level co-occurrence matrix, respectively. After that, the Gaussian process regression is exploited for training the prediction model from the extracted features to subjective opinion score. Also, to verify the generalization ability of the existing IQA metrics, we establish a real-world underwater IQA dataset with subjective scores. Extensive experiments conducted on public benchmark datasets and our constructed dataset both demonstrate that our proposed MFEM achieves better prediction performance comparing with several state-of-the-art IQA metrics. The code and dataset are available at: https://github.com/Hou-Guojia/MFEM.