Video online stabilization method and system

Abstract

The embodiment of the invention provides a video online stabilization method and system. The method comprises the steps of obtaining a to-be-tested video and a trained convolutional neural network; inputting a current frame of image and a plurality of previous frames of images in the to-be-tested video into the trained convolutional neural network to obtain deformation parameters; converting the next frame of image of the current frame of image through the deformation parameter, obtaining the stabilized next frame of image, and taking the stabilized next frame of image as the current frame ofimage again; and until all frames of the to-be-tested video are processed, obtaining the stabilized to-be-tested video. According to the video online stabilization method and system provided by the embodiment of the invention, the problem of video stabilization is solved through the deep convolutional neural network, the degree of parallelism of the embodiment of the invention is high, the operation speed on a GPU is ten times faster than that of a traditional offline method, and the application scene of video stabilization is expanded.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of image processing, and in particular to a video online stabilization method and system. Background technique [0002] With the rapid development of mobile devices and social networks, shooting with mobile devices such as mobile phones has become one of the common forms of online media. Video stabilization technology is essential for most hand-held videos due to high-frequency jitter. Several 2D, 2.5D and 3D based stabilization techniques have been proposed previously, but to our knowledge no deep convolutional neural network based solution has been proposed so far. The main reason for this omission is the lack of training data and the challenge of modeling the problem using neural networks. [0003] In the field of video stabilization, Grundmann et al. proposed to use L-based 1 The path optimization method stabilizes the hand-held video. In 2013, Liu et al. proposed to divide the v...

AI Technical Summary

Problems solved by technology

There are also many applications that directly generate video, such as scene dynamic generation, frame interpolation and deblurring, but predicting long video sequences is still a challenging problem, so all the above works only use two or a few consecutive frames as training samples

Images