Quality Improvement Of Large Motion Video Int-Erpolation Based On Motion Estimation And Str-Uctural Feature

Video interpolation is a technique that synthesizes frames that do not exist between frames to improve the visual effect of the video.

We made a study to improve the quality of interpolation frame by making large motion video interpolation based on structural features and motion estimation. The proposed architecture consists of an end-to-end learning network. This method uses generalized Warping model by motion estimation and boundary-based structural features to better reflect the structural information in large motion video interpolation. It also helps to further improve the quality of interpolation frames through post processing networks.

The proposed video frame interpolation framework consists of two main components, as shown in the figure.

Fig. A survey of the proposed video frame interpolation network architecture

In our approach, we use BDCN (Bidirectional cascade network) to obtain the boundary image from the input frames I_t-1(x) and I_t+1(x) and input the 4-channel image into the feature extraction network. Meanwhile, we obtain the contextual information C_t-1 and C_t+1 of the input frame using the context extraction network. Then, the feature images of the obtained input frames are input into submodules to obtain the flow vector, based on which we obtain the waffing results for the input frame and context images, respectively.

Then, we input the warping results for the intermediate frame and context image obtained by occlusion process on two warping images for the input frame into the redundant processing network to obtain redundant information and combine with the intermediate frame to generate the final interpolation frame.

We have verified the excellent performance of the proposed method on several public benchmark datasets. Quantitative and visual evaluations show that the proposed method outperforms previous methods on a variety of datasets. Experimental results show that the proposed method can improve the quality of video interpolation frame even for occluded regions.

The results were published in the "Information Technology" under the title of "Quality improvment of large motion video interpolation based on motion estimation and structureal feature"(https://www.elibrary.ru/item.asp?id= 25795345).