Dynamic texture synthesis method based on double-identification flow convolutional network

Abstract

The invention provides a dynamic texture synthesis method based on the double-identification flow convolutional network. The method mainly comprises steps of design of an apparent flow in a texture model, design of a dynamic flow in the texture model, new texture generation, dynamic texture synthesis and texture style conversion. The method is characterized in that target identification and optical flow prediction are accomplished based on the pre-training convolutional network (ConvNet), after given dynamic state texture input, each appearance of input textures is packaged according to filter data responded by a target identification convolutional network, modeling of a dynamic state of the input textures is carried out according to filter data responded by an optical flow convolutional network. For generating new textures, optimization of a noise input sequence is required to match characteristics of each identification flow in the input textures, and appearance of one texture and a dynamic state of another texture are combined to generate a brand new dynamic texture. The method is advantaged in that a high quality sample matched with pixel-by-pixel appearance and time-varying conditions of the input textures can be generated.

Description

technical field [0001] The invention relates to the technical field of dynamic texture synthesis in computer graphics, in particular to a dynamic texture synthesis method based on a double recognition flow convolution network. Background technique [0002] Dynamic textures refer to image sequences with time-related repetitive features that describe a certain dynamic landscape. They exist widely in nature, such as ocean waves, waterfalls, flying flags, flying birds, and so on. Many scholars at home and abroad have done a lot of work on dynamic texture synthesis by using simulation methods based on physical models, and achieved good results in the synthesis of some phenomena. The synthesis method based on physical model simulation analyzes the physical laws of specific phenomena, establishes a simplified physical model, and performs illumination calculation and drawing. Due to the complexity of dynamic landscapes, these physical models are difficult to construct, computationa...

AI Technical Summary

Problems solved by technology

Due to the complexity of dynamic landscapes, these physical models are difficult to construct, computationally complex and only applicable to specific phenomena

Images