Color constancy method based on cascade fusion feature confidence coefficient weighting

Abstract

The invention discloses a color constancy method based on cascade fusion feature confidence weighting, which provides stable color features for unmanned driving, underwater object recognition, three-dimensional object reconstruction and other computer vision tasks, and comprises the following steps: (1) shooting images and videos under a natural scene light source, and making a data set applied to the color constancy method; (2) according to the particularity of light source colors, weighting a network structure based on cascade fusion feature confidence; (3) performing two-stage training on the network structure by using the data set; and (4) removing the estimated scene light source from the image or the video to realize color constancy of the image and the video. According to the method, the shallow edge texture features and the deep fine-grained deep features in the image are fused in a cascading mode, the feature estimation light source capable of providing more information for light source estimation in the image is fully utilized, and the problem that the light source estimation accuracy is low when a current color constancy method faces a complex environment is solved; and the accuracy of the color constancy method and the robustness of the method in a complex environment are improved.

Description

technical field [0001] The invention relates to a method for restoring an image or video to the color under a standard light source using deep learning technology, which belongs to the fields of artificial intelligence, computer vision and image processing, and specifically relates to a color constancy method based on cascade fusion feature confidence weighting The design is mainly used in various tasks of target detection and recognition in computer vision, especially in the color correction of underwater images and videos and target detection and recognition in the field of unmanned driving. Background technique [0002] As one of the most basic and direct features of visual information, color has been widely used in the fields of image processing and computer vision. The image and video imaging process is affected by many aspects such as scene illumination, reflectivity of the object surface, and the response function of the imaging sensor. The color reflected by the sur...

AI Technical Summary

Problems solved by technology

[0004] Most of the convolutional network structures currently applied to the color constancy method are improved from the network structure of classification or recognition tasks. Although the convolutional network structure applied to the color constancy method further improves the accuracy of light source estimation, due to The particularity of the color of the light source, the color constancy method needs to apply local features in the image that can provide more information for the light source estimation light source, and the current color constancy method treats all the extracted feature information equally and cannot make full use of it. Features that can provide more information for illuminant estimation Accurately estimate the illuminant

At the same time, the network structure currently applied to the color constancy method uses the fine-grained feature information in the image to estimate the light source by deepening the network layer, ignoring the importance of the shallow edge texture feature information in the image to the light source estimation, resulting in the application of color constancy. Although the network structure of the property method has powerful feature extraction capabilities, it is not sensitive to changes in light source conditions, has light source invariance, and cannot make accurate estimates of light sources.

Images