Published in Computational and Systems Neuroscience (CoSyNe), Feb 2016.
Models for perceptual image distortion are generally designed to capture properties of human vision, such as the spatial frequency dependence of contrast sensitivity, and the masking effects of superimposed oriented patterns. However, the widely-used Structural Similarity (SSIM) index (Wang et al., 2004), is based on a unique construction devised to disregard changes in local luminance or contrast, while emphasizing changes in local structure. Since these properties are evident in the responses of early visual neurons, we wondered whether a more explicit model of physiological responses might provide a more suitable substrate for constructing a distortion measure. We built a functional model of early spatial visual processing (LGN- GC) that incorporates known properties of retina and LGN. Similar to the model of Mante et al., 2008, the model includes bandpass linear filtering, rectification, and local luminance and contrast gain controls. The distortion between an original and corrupted image is determined by passing each through the model, and measuring the Euclidean distance between the two response vectors. The model parameters (filter sizes and amplitudes) were fit to a database of human perceptual quality judgments (TID2008 - Ponomarenko et al., 2009). We find that the fitted parameters are consistent with measured physiological properties of LGN neurons in the macaque monkey. Moreover, LGN-GC outperforms SSIM, and performs comparably to multi-scale SSIM (MS-SSIM) at predicting perceptual distortions (despite its restriction to a single spatial scale), explaining more of the (cross-validated) variance in the human data (SSIM: 60%, MS-SSIM: 64%, LGN-GC: 67%). Finally, we performed a direct comparison of LGN-GC to SSIM by examining stimuli optimized to differentiate them (known as MAD competition; Wang et al., 2008). This comparison provides evidence that distance as measured by our physiologically-inspired model corresponds more closely with human perception than SSIM.