The Importance Of Phase To Texture Similarity

Although the importance of the Fourier phase to image perception has been addressed, it is unknown whether this is the case for texture similarity or not. Based on psychophysical experiments, we first show that the phase data is more important to human visual perception of texture similarity than the magnitude data. We further examine the ability of a total of 51 computational feature sets on exploiting the phase data for texture similarity estimation. However, it is found that for these feature sets the magnitude data is more important than phase. Since it has been shown in early work that there is inconsistency between the similarity data derived from human observers and the 51 feature sets, we attribute this outcome (magnitude/phase importance) to the difference between the manners in which humans and the feature sets exploit the phase data. Therefore, we are motivated to enable the 51 feature sets to exploit the phase data for effective estimation of texture similarity. This is achieved by fusing the features extracted from the original and phase-only images. It is shown that this type of fused feature sets yield better results than those derived using the 51 original feature sets. In particular, we show that this finding can also be propagated to convolutional neural network features. We believe that the improved results should be attributed to the importance of phase to texture similarity.