关键词: 动态纹理, 合成, 自纠正, 图像质量评价, 回归优化

Abstract: Dynamic texture is one of the dynamic models in computer vision. It has statistical stationarity in the spatial range and random repetition in the time dimension. The goal of dynamic texture synthesis is to generate an image that is visually similar to a given texture. When performing dynamic texture synthesis, the accumulation of regression prediction errors is a key issue of leading to the degradation of texture quality. Therefore, this paper proposes a dynamic texture synthesis model based on self-correction mechanism. The model uses indicators such as clarity, structural similarity, and optical flow to determine the optimal data range and finds the optimal extreme point. Through the self-correction mechanism, the original data is replaced with optimized data, and the optimized data is used for regression prediction. Finally, a convolutional autoencoder is used to reconstruct the prediction data into high-dimensional dynamic texture video frames. Experiments  are conducted on the DynTex database and the model proposed in this paper is compared with several typical dynamic texture synthesis models. The experimental results show that, compared with other models, the Mean Square Error (MSE) value calculated by the dynamic texture video frame and the real video frame synthesized by this model is smaller, and the values of Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are larger. It solves the problems of residual image, blur, and noise in dynamic texture synthesis, so as to generate a better visual effect and longer dynamic texture sequence. At the same time, the effectiveness of the proposed modeling method is verified.

Key words: dynamic texture, synthesis, self-correction, image quality evaluation, regression optimization