Research on Color Constancy Algorithm of Weighted Fully Convolution Neural Network based on Multi-Path Feature Fusion

Abstract

Color constancy refers to the ability to make the color of an object not change with the change of the light color of the scene. The use of convolutional neural networks improves the accuracy and stability of color constancy. But the traditional convolutional neural network structure cannot apply the semantic features in the image to the color constancy, which makes the illumination estimation fuzzy, greatly reduces the quality of network training. In order to solve this problem, in this work we propose a new network structure called multi-path feature fusion weighted fully convolutional neural network structure. The network structure contains three convolutional neural network branches for fully extracting semantic features in images. After that, the extracted features are fused and weighted to accurately estimate the light source. Experimental results show that the multi-path feature fusion weighted network structure can make fully use of the semantic feature information in the image to accurately estimate the scene light source. At the same time, it is a more lightweight network, which achieves the simultaneous improvement of the accuracy and speed of light source estimation, meets the requirements of mobile devices for lightweight network models.