On the perception of band-limited phase distortion in natural scenes

It is widely believed that the phase spectrum of an image contributes much more to the image’s visual appearance than the magnitude spectrum. Several researchers have also shown that this phase information can be computed indirectly from local magnitude information, a theory which is consistent with the physiological evidence that complex cells respond to local magnitude (and are insensitive to local phase). Recent studies have shown that tasks such as image recognition and categorization can be performed using only local magnitude information. These findings suggest that the human visual system (HVS) uses local magnitude to infer global phase (image-wide phase spectrum) and thereby determine the image’s appearance. However, from a signalprocessing perspective, both local magnitude and local phase are related to global phase. Moreover, in terms of image quality, distorting the local phase can result in a severely degraded image. These latter facts suggest that the HVS uses both local magnitude and local phase to determine an image’s appearance. We conducted an experiment to quantify the contributions of local magnitude and local phase toward image appearance as a function of spatial frequency. Hybrid images were created via a complex wavelet transform in which the the low frequency magnitude, low frequency phase, high frequency magnitude, and high frequency phase were taken from 2-4 different images. Subjects were then asked to rate how much each of the 2-4 images contributed to the the appearance of the hybrid image. We found that local magnitude is indeed an important factor for image appearance; however, local phase can play an equally important role, and in some cases, local phase can dominate the image’s appearance. We discuss the implication of these results in terms of image quality and visual coding.