Dynamic Texture Synthesis

Dynamic textures are representations of such textured surfaces with repetitive, time-varying visual patterns which form an image sequence that exhibits certain stationary properties in time. These include sea-waves, smoke, foliage, whirlwind, dense crowds, and traffic scenes. The aim of dynamic texture synthesis in our work is to learn some model which best represents the given sample textures, and then using the learnt model to predict and extrapolating the finite dynamic textures to an infinite one sharing similar dynamic behaviors with the original input.We first propose the linear dynamic system (LDS) model from the gray level video sequence to colored dynamic textures by compressing the color channels in the YUV color space using laplacian pyramid and radial basis functions (RBF). The dynamic color texture synthesis model we propose is memory efficient and has the ability to better capture complex dynamic behaviors. Further, we develop the multi-resolution auto-regressive moving average model (MARMA) based on the first order linear dynamic system by transforming the dynamic textures into different frequency scales using multi-resolution analysis descriptors. We perform the comparatively study using three representative multi-resolution models: Laplacian pyramid, haar wavelet and the steerable pyramid. The multi-resolution analysis method forms a multi-level LDS model which not only improves the synthesis algorithm, but also shed a light on dynamic texture comparison.In summary, the dynamic color texture synthesis model we propose can deal with large size video sequence. Especially for the RBF based model, it is not only memory efficient, but also capturing the structural information better than the previous RGB color approach for the complex dynamic color textures. The multi-resolution auto-regressive moving average model can capture the local and global dynamic behaviors in different scales, thus causes a better synthesized result than the LDS model. Our experiments show that steerable pyramid performs better than other multi-resolution analysis methods in our MARMA model.