PURPOSE: Procedural textures are normally generated from mathematical models and commonly used in computer graphics applications such as games and animations. Classification of procedural texture is very important for selecting proper generation models. Pioneering study on human perception of procedural texture was conducted on perceptual features and semantic attributes. In a psychophysical experiment, Liu et al. presented textures to human observers and asked them to rate textures on twelve perceptual features. Dong et al. further obtained 43 semantic attributes on the same dataset. In this study, we proposed a two-view learning method on procedural textures classification using perceptual features and semantic attributes.
METHODS: Texture perceptual features such as coarseness, density and direction, represents the subjective perceptions of the observers, can be seen as a special type of abstract information. While texture semantic attributes are more concrete descriptions of texture images given by observers, such as honeycombed, lined and netlike. The perceptual features and binary semantic attributes of 450 textures were selected in this experiment. We use canonical correlation analysis as the basic method of two-view learning to combine the perceptual features and semantic attributes to obtain a single feature vector, which is more discriminative than any of its single feature counterpart. Given a texture, the two properties will be simultaneously considered for texture classification.
RESULTS: The 450 textures are generated from 23 representative procedural texture generation models. We tested the classification performance with perceptual features and semantic attributes as features respectively. The accuracies of cross validation for the two properties was 60.49% and 67.09% respectively. Then we tested the classification performance with the combined features and the accuracy was 79.01%, which outperformed the 12 perceptual features and 43 semantic attributes by more than 18% and 11% respectively.
CONCLUSIONS: Experimental results demonstrated that both perceptual features and semantic attributes are effective properties in human perception, using the two-view learning method to combine the two properties is more effective in procedural textures classification.