Identifying Perceptual Features of Procedural Textures J Liu1, J Dong1, L Qi1, M Chantler2 |
|---|
1Department of Computer Science and Technology, Ocean University of China, China |
| Identifying perceptual texture features is important for texture generation, browsing and retrieval. This work focused on investigating perceptual features of procedural textures. We generated 450 samples using 23 generation methods. We designed two psychophysical experiments: free grouping and rating. First, twenty observers were asked to group the 450 samples, from which a similarity matrix of 23 methods was created. Hierarchical cluster analysis (HCA) was applied to the matrix and these methods were clustered into 10 classes. Second, observers rated each sample six times in the 12 texture description dimensions proposed by [Rao A.R and Lohse G.,1996, Vision Research, 36(11), 1649-1669] using 9-point Likert scales. We trained a support vector machine model for prediction based on the HCA results with the 12-dimensional features. For all texture generation methods, the accuracy for predicting a given sample belonging to a certain class was 59.22% for the leave-one-out test. However, when we selected five near-regular texture generation methods based on the rating data, the prediction accuracy was raised to 91%. These results indicated that the 12 perceptual features could be used to discriminate near-regular texture classes as human perceived; however, they are not good enough for discriminating random textures classes. [NSFC Project No. 61271405] |
| Up Home |
|---|