Intrinsic image decomposition method and device based on deep learning

Abstract

The invention discloses an intrinsic image decomposition method and device based on deep learning. The method comprises that a 3D model is selected from multiple 3D models and loaded with a physics based renderer, illumination is set randomly, an image is rendered at a random visual angle, the renderer obtains corresponding reflection and illumination components, and above operations are repeatedto generate a large batch of data sets with marked intrinsic image decomposition; a full convolutional neural network is trained into an intrinsic image decomposition network via the generated data sets; and the intrinsic image decomposition network is used, and a decomposition target of expected output is obtained from a decomposition result of predicted output. Via the intrinsic image decomposition method, a large batch of marked data sets is obtained in an image rendering manner, the deep neural network is trained to obtain a decomposition model of high robustness, and the loss network is used to further improve the generalization performance and avoid difficulty in design of the loss function.

Description

technical field [0001] The present invention relates to the technical field of eigenmap decomposition, in particular to a method and device for eigenimage decomposition based on deep learning. Background technique [0002] The existing eigengraph decomposition techniques mainly fall into the following categories: [0003] 1) Numerical optimization method based on prior assumptions [1,2]. Such methods rely on the assumption of continuity of some physical properties or phenomena such as lighting, object surface, depth, etc. For example, [1] constructs the corresponding energy function and constraints by relying on the smoothness assumption of direct and indirect irradiance, and optimizes the minimum square error; [2] learns the geometry of the object surface and the prior distribution of illumination through the Gaussian mixture model, According to the corresponding observations, the loss functions corresponding to different components are constructed and weighted for optimi...

AI Technical Summary

Problems solved by technology

Many numerical optimization methods often require lengthy and time-consuming optimization for each input, which limits their application in the real-time field

At the same time, the corresponding observations and assumptions have their limitations and are only valid in specific circumstances

However, learning-based methods are limited by the difficulty of labeling datasets, and it is often difficult to obtain a large number of densely labeled image datasets.

Images