Research On Color Constancy Of Image

As one of the research hot spots in the field of computer vision,the research on color constancy of image is of great significance for industrial inspection,object identification,unmanned driving and some special fields.When the external environment changes,imaging equipment may be unable to accurately identify the color of the scene,resulting in inaccurate image information during imaging,or even unable to distinguish the object in the image.In response to this problem,this article is dedicated to research on color constancy of image.By reducing the influence of external environmental factors on the imaging image,a stable color feature that can describe the object itself is obtained,so that the image can get accurate color restoration.1.This article realizes traditional illumination estimation algorithms.Based on the correction result traditional illumination estimation algorithm by comparing the Color Checker data set,and then provide help for the improvement of the gray world algorithm in the next chapter.2.Aiming at the deterioration of correction in traditional illumination estimation algorithm,the gray world,an improved color constancy method based on image entropy is proposed.This method calculates the entropy value of the image to determine the amount of color information contained in the image,and solves the calculation error caused by the smaller amount of image color information.Experimenting with this algorithm on the Color Checker data set,and comparing the results,it is found that the improved algorithm error has decreased in the two algorithm evaluation indicators of Euclidean distance and angle error,which plays a positive role in improving image quality.3.In order to improve the applicable range of the algorithm,this paper designs a color constancy method based on deep residual network to estimate illumination.This method improves the original residual network: The capsule model is introduced to replace the convolution layer in the original residual module,and encapsulate the convolutional neurons;the attention module is introduced to weight the image;the Leaky Re LU activation function is introduced to avoid the death of some neurons.This method actively extracts features without having to base on various assumptions.The results show that compared with traditional methods,the output results obtained by the network model of this paper have a decrease in the errors of the median and mean,and the corrected image is improved.