Optimized and Efficient Color Prediction Algorithms Using Mask R-CNN

Abstract

Color cognizant capability has a significant impact in service robots for object detection based on color, traffic signal interpretation for autonomous vehicles, etc. Conventional clustering algorithms such as K-means and mean shift can be used for predicting the dominant color of an image by mapping the pixels from RGB to HSV and clustering them based on HSV values, thereby picking the cluster with the most pixels as the dominant color of the image, but these approaches are not solely dedicated to the same outcome. This research’s goal is to introduce novel techniques for predicting the dominant color of objects in images, as well as pixel extraction concepts, which allow these algorithms to be more time and efficiency optimized. This investigation appraises propriety of integrating object detection and color prediction algorithms. We introduce a dominant color prediction color map model and two new algorithms: average windowing and pixel skip. To predict objects in an image prior to color prediction, we combined the Mask R-CNN framework with our proposed techniques. Verification of our approach is done by creating a benchmark dataset of 200 images and comparing color predicted by algorithms with actual color. The accuracy and runtime of existing techniques are compared with those of the proposed algorithms to prove the superiority of our algorithms. The viability of the proposed algorithms was demonstrated by scores of 95.4% accuracy and color prediction time of 9.2 s for the PXS algorithm and corresponding values of 93.6% and 6.5 s for the AVW algorithm.