Procedural Textur Generation Based On Perceptual Consistency

Texture generation and representation based on procedural texture models is an important research field. Research output has been widely used in computer games, animation, geology, artistic design and many other fields. An advantage of procedural texture generation models is that different textures can be efficiently produced by varying a set of input parameters to mathematical models. However, one challenging problem is how to select a procedural model and corresponding paramenters to generate certain texture that is consistent with descriptions based on human perception. This paper introduces a framework to automatically determine procedural models and corresponding parameters to generate new texture with varied perceptual scales, while preserving appearance similar to the example. The work is introduced in detail as follows:1. Construction of procedural textures database. The database includes four hundred -fifty hightmaps generated by twenty-three procedural models and the corresponding natural-like textures rendered under the same lighting and view conditions. Two psychophysical experiments where conducted in order to capture the human’s judgements of textures similarity and perceptual values for 12 features.2. Identifying perceptual features of procedural texture models and revelent perceptual dimensions of procedural textures. Perceptual salient characteristics of different procedural models are identified by statistical analysis. Correlations between perceptual features are identified by correlation analysis. Based on the construction of perceptual texture space, three relevent perceptual dimensions are identified. Experimental results show that the weights of different perceptual dimensions varied with stimuli, while the features combinations corresponding to individual dimensions remained unchanged. Moreover, similarity measure in the PTS are more consistent with human’s perception.3. Proposing a PCA-based convolutional network (PCN) for unsupervised feature learning. The architecture of PCN is composed of cascaded feature extraction stages and a nonliear output stage. In the convolutional layer, the filters are simply learned by PCA. For the higher convolutional layers, the filter banks are learned from the combinations of feature maps that are obtained in the lower layers. Moreover, sampling the patches with a stride and pooling layer dramatically reduce the dimensionality of data. The experimental results show that features learned from PCN are robust in various tasks. It can achieve comparative or even better performance than state-of-art methods but is much more efficient. Then, PCN is used to extract features from given texture, and based on the computational features, scales of perceptual features and corresponding procedural model can be accurately predicted.4. Proposing a framework for perceptually-based procedural texture generation from examples. Based on the computational features extracted from input texture, the perceptual scales and a procedural model can be predicted. The parameters of the model are determined by performing perceptually similarity measurement with a certain set of points in the perceptual texture space. Thus, we can generate a new texture with varied perceptual scales while preserving appearance similar to the example. The experimental results verify the effectiveness of the proposed method. The method supports a large variety of procedural texture models with a unified approach.