Color Constancy Calculation Based On Image High-Level Semantics

Color constancy is a kind of psychological perception produced by the human eyes when observing external things.It can keep the human eye's stability of real color perception on the surface of the object under the condition of changing ambient light source conditions.In the field of artificial intelligence,color constancy calculations are closely related to many computer vision tasks,in order to achieve correction of the color of the light source and to present images that are consistent with human perception of the world.Most of the existing classic color constancy algorithms are based on the low-level features of the image to carry out the calculation,the psychological perception of the human being while observing the outside world and the understanding of the semantics of the deep consciousness are not considered.Therefore,the performance of such algorithms is often unstable and the scope of application is also not wide.Considering the human perception and the understanding of image semantic content,establishing the scene semantic classification and the object semantic classification set based on the high-level semantics,and optimizing the existing algorithm by using the classification set,finally realizing the color constancy calculation in this paper.The paper proposes two color constancy calculation methods based on the highlevel semantics of images: The first is the color constancy algorithm based on image scene semantics: According to the Grey Edge color constancy algorithm framework,the characteristics and disadvantages of Grey World algorithm,White Patch algorithm,Shades of Grey algorithm,general-Grey World algorithm,1st-Order Grey Edge algorithm and 2nd-Order Grey Edge algorithm are summarized,then establishing the indoor/outdoor scene semantic data set based on the original color constancy database,and use the pattern in different classification scenarios in the data set.The Pattern Search Method(PSM)optimizes the above algorithm parameters,and finally shows the effectiveness and stability of the method through objective comparison evaluation.The second is the color constant algorithm based on the semantic classification of image objects: using the Mask-RCNN deep learning model to realize the object semantics segmentation,and the semantic classification sets of image objects are established based on semantic themes.Based on this analysis,the Grey World algorithm is improved.The final evaluation results show that the algorithm is superior to the existing algorithms.