Synthesizing and Recovering Dynamic Patterns Using the Causal Reasoning Model

A new method using the causal reasoning ConvNet mode is proposed for synthesizing and recovering dynamic patterns. In the model which defines the probability distribution, a multi-layer filter is used to capture the spatiotemporal information. The problem of synthesizing and recovering pattern design is to capture the optimal synthesis or recovery based on the training model and the solution to maximize the objective function. In the algorithm, the synthetic video generated by the ConvNet of opponent training is matched with the observed (recovered) video frame, and the multi-layer parameters of the filter are updated. The maximum likelihood algorithm is designed to get the synthesis (recovery) of the real video patterns. The DynTex database is used to evaluate the effectiveness of this method and the results are compared with typical synthesis methods.