A color constancy method based on convolutional autoencoders

Abstract

The invention discloses a color constancy method based on a convolutional self-encoder. Firstly, an image without color shift is obtained, and an image data set is made according to the image without color shift; then, the convolutional self-encoder is trained through the image data set. Establish a mapping network from the image with color shift to the image without color shift; finally realize the color constancy of the image through the mapping network. The invention can achieve good color constancy effect, and can effectively extract the hidden light source color information in the original image by utilizing the powerful encoding ability of the convolutional self-encoder, thereby performing image color correction. The present invention is tested on multiple international common color constancy databases, and the result proves that the method of the present invention can obtain very good effect of color constancy under the condition of using less parameters.

Description

technical field [0001] The invention belongs to the technical field of computer vision and image processing, and in particular relates to the design of a color constancy method based on a convolutional self-encoder. Background technique [0002] Color constancy, as a kind of perceptual constancy, is a very important function in the human visual system. It can help us maintain a stable perception of objects in the scene under changing light sources. Achieving good color constancy in computer systems is an important research direction in the field of computer vision, and has important practical significance for its downstream tasks such as image enhancement, denoising, and recognition. The traditional color constancy method is based on some simple physical scene assumptions or machine learning methods, and has the characteristics of simple algorithm, easy implementation, and strong adaptability, but the effect is relatively poor. In recent years, with the rapid development of...

AI Technical Summary

Problems solved by technology

Convolutional autoencoders are widely used in image denoising, signal compression, style transfer and other fields due to their powerful representation learning capabilities. However, in practical applications, as the number of network layers deepens, a large number of convolution and pooling operations The effective information of the image is continuously lost, which leads to the "information bottleneck" problem, so that it is difficult for the decoder to decode and reconstruct the original input, and the ideal image processing effect cannot be obtained.

Images