Learning perceptual texture similarity and relative attributes from computational features

Abstract

Previous work has shown that perceptual texture similarity and relative attributes cannot be well described by computational features. In this paper, we propose to predict human's visual perception of texture images by learning a nonlinear mapping from computational feature space to perceptual space. Hand-crafted features and deep features, which were successfully applied in texture classification tasks, were extracted and used to train Random Forest and rankSVM models against perceptual data from psychophysical experiments. Three texture datasets were used to test our proposed method and the experiments show that the predictions of such learnt models are in high correlation with human's results.

A Perception-Inspired Deep Learning Framework for Predicting Perceptual Texture Similarity

“…This further proves that the perceptual similarity predicted by the similarity network can fit the psychophysical data well. We compare our method with the similarity regression method proposed in [16] and [55], the results are shown in Table. I. As can be seen from Table.…”

“…I, our method produces better results with smaller Mean Squared Errors (MSE) and higher correlation coefficients. It should be noted that the methods reported in [16] and [55] require to select the best features but ours does not need this stage.…”

“…The optimization converges at 340,000 iterations, and the MSE on the test set is 0.013. We also compare our method with the similarity regression method proposed in [16] and [55], and the results are shown in Table. II.…”

“…The most related work to ours is by Lou et al [16], in which Random Forests were employed to regress the similarity data and produced promising results. In their work, after computational features have been extracted from the sample textures, the paired texture features were combined together and then sent to the random forests classifier for the prediction of fine-grained similarity values.…”

“…However, Dong et al demonstrated that there is no simple relationship between the perceptual attributes and the computational features of texture images [13]. Although computational features perform well in many tasks, such as texture classification [14] and semantic segmentation [15], they still produced large deviation when used to estimate perceptual texture similarity [16]. On the other hand, researchers attempted to predict texture similarity by constructing a perceptual texture space using manifold learning [12], [17], [18].…”