【摘要】 针对现有的仅可察觉差异(just noticeable difference,JND)模型未能考虑水下图像特有的颜色选择性衰减的问题,提出一种基于水下图像的JND预测模型.该模型考虑光在水下特殊的吸收和散射特性,设计水下图像特征提取模块,建立基于特征提取模块的全参考水下图像损伤感知网络,并将其命名为感知有损/无损预测器,用于预测畸变图像相对于参考图像而言是否有感知上的损失.进一步采取基于反馈的搜索策略,根据预测器的结果自适应调整图像压缩程度并搜索JND图像.实验结果表明,与现有的JND模型相比,在峰值信噪比等多个指标下,所预测的JND图像与实际真值之间的误差最小.更多还原

【Abstract】 None of the existing just noticeable difference(JND)models specifically consider the phenomenon of selective color attenuation unique to underwater images.This paper proposes a JND prediction model based on underwater images.In this model,special absorption and scattering characteristics of light underwater are considered,and an underwater image feature extraction module is designed.A full-reference underwater image lossy/lossless perception network based on the feature extraction module is established,and it is named the perceptually lossy/lossless predictor,which is used to predict whether the distorted image has endured a perceptual loss relative to the reference image.Then we adopt a feedback-based search strategy and adaptively adjust the degree of image compression according to the results of the predictor and search for JND images.Experimental results show that,compared with the existing JND model under multiple indicators such as peak signal-to-noise ratio(PSNR),the minimal error between the predicted JND image and the ground truth is found.更多还原