【摘要】 本文针对能源大数据可视化仪表板配置过程中采用未经用户感知设计的彩虹调色板易于造成设计效率低且视觉效果难以专业化等问题，提出一种基于深度学习的数据可视化仪表板着色推荐方法。首先使用目标检测算法Grounding Dino实现对仪表板中图表元素的准确识别与定位；其次，使用区域分割法模型（segment anything model,SAM）对识别的图表进行精准分割，从而更好地捕捉图表的颜色信息；最后采用K-means聚类算法对分割后的图表进行主颜色提取和调色板推荐，从而减少仪表板中彩色文字等对主色造成的影响。实验结果表明，该模型的颜色提取的平均覆盖率较卷积神经网络（convolutional neural network,CNN）、生成对抗网络（generative adversarial nets,GANs）和半自动着色算法分别提高20.68%、10.76%和32.04%，覆盖率的波动分别降低4.62%、0.73%和29.54%，生成的着色方案与仪表板背景色的平均色差比美工人员调色板低2.35，能够有效提高仪表板颜色的生成质量和协调度，协助设计人员快速完成可视化仪表板的自动着色过程。更多还原

【Abstract】 A deep learning based shading recommendation method for data visualization dashboards is proposed to address the issues of low design efficiency and difficulty in professionalization of visual effects caused by using rainbow color palettes that have not been perceived by users during the configuration process. Firstly,the Grounding Dino object detection algorithm is used to accurately identify and locate chart elements in the dashboard; Secondly,the segment anything model(SAM) is used to accurately segment the recognized charts,thereby better capturing the color information of the charts; Finally,the K-means clustering algorithm is used to extract the main colors and recommend color palettes for the segmented charts,thereby reducing the impact of colored text and other factors on the main colors in the dashboard. The experimental results show that the average coverage of color extraction in this model is improved by 20.68%,10.76%,and 32.04% compared to CNN(convolutional neural network),GANs(generative adversarial nets),and semi-automatic coloring algorithms,respectively. The fluctuation of coverage is reduced by 4.62%,0.73%,and 29.54%,respectively. The average color difference between the generated coloring scheme and the dashboard background color is 2.35% lower than that of the artist’s palette,which can effectively improve the quality and coordination of dashboard color generation and assist designers in quickly completing the automatic coloring process of visualized dashboards.更多还原