图像的恰可察觉失真(just noticeable distortion, JND)阈值是指人眼能够察觉的最小失真, 通常被用于去除图像/视频压缩中的视觉冗余. 针对 JND 模型对颜色和结构特征利用不充分的问题, 提出了一种基于颜色复杂度和结构张量的 JND 模型. 首先, 计算图像的颜色复杂度, 将其转换为与视觉敏感度相关的权值, 和对比掩蔽模型结合以提升模型的准确性; 然后, 利用结构张量对局部特征进行表示, 建立基于局部结构特征的调制因子, 估计结构不规则区域的视觉冗余程度; 最后, 将基于颜色复杂度的 JND 模型和基于结构张量的调制因子结合, 建立基于颜色复杂度和结构张量的 JND(complexity structure tensor based JND, CSJND)模型. 实验结果表明, 相比于已有的模型, 该模型在主观感知质量相同的前提下, 能使 PSNR 值明显降低; 该模型更加符合人眼的视觉特性, 能更准确地估计出 JND 阈值.

The just noticeable distortion (JND) threshold refers to the minimum distortion at which eyes can perceive. JND can be used to remove visual redundancy derived from image or video compression. Considering that JND models do not make full use of colour features and structural information, this study proposes a JND model based on colour complexity and structure tensor. First, the colour complexity is estimated and it is used to calculate the visual weight values related to the sensitivity of human eyes. Then the estimated colour complexity is combined with the contrasting masking effect to improve the accuracy of the model. Next, utilising the local structure tensor to represent local features, the modulation factor is established to calculate the visual redundancy of irregular regions. Finally, the colour complexity structure tensor based JND (CSJND) model is estimated by combining the colour-complexity-based JND model and structure tensor modulation factor. Experimental results show that the proposed CSJND model can acquire a noticeably lower peak-signal-to-noise ratio as compared with some existing JND models while also achieving the same subjective perceptual quality. This is more consistent with human visual perception. The proposed CSJND model can also calculate the JND thresholds more accurately.