A Novel Color Difference-Based Method for Palette Extraction and Evaluation Using Images of Birds

Color palettes are important sources of inspiration for designers. In most coloring studies, designers and interior architects use ready-made color palettes that are known to be harmonious and popular with clients. In this study, a new color palette extraction and evaluation method is proposed that differs from existing methods because it accelerates designers’ product coloring processes by identifying the color-area and color-neighborhood relationship. Furthermore, attractive color palettes will be identified by focusing on living creatures, primarily colorful birds. For this purpose, a new data set of images of birds was created. In the literature, the success of color palettes is usually evaluated through questionnaires, while measurements are made without relying on numerical metrics. In this study and, for the first time, a new metric based on CIEDE2000 and the Hungarian algorithm to measure the accuracy of color palettes is demonstrated. In the proposed method birds are first segmented from background using Mask R-CNN and PointRend models. Then, Fuzzy c-means (FCM), Gaussian Mixture Model (GMM), k-means, Mean Shift, and mini batch k-means algorithms were used to extract the color palette of each bird image. Color reduction based on the CIEDE2000 was performed on the color palettes using various threshold values, resulting in the final color palettes. In the last step, color-area and color-neighborhood relationships are shown through graph-based color palettes. Experimental results showed that the PointRend method, with a 93.92% success rate, produced the most successful results. However, k-means and GMM methods were more successful in extracting the colors of birds. The evaluation of the palettes was performed with the proposed numerical metric rather than questionnaires, resulting in more robust and repeatable findings. Furthermore, the proposed graph-based color palette results in considerably simplified coloring process by reducing the computational complexity for an industrial product.