Segmentation-Guided Dominant Color Classification of Objects in Drone and Ground Imagery Using Masked K-Means Clustering and Hue-Saturation-Value Rules

Abstract

Identifying the dominant color of objects seen from a drone is useful for surveillance, tracking, and logistics, but color estimates taken from rectangular bounding-box crops are biased toward the background around the object. This study presents a color detection pipeline that combines a pretrained instance segmentation network, K-means clustering, and a 16-class rule-based classifier in the hue-saturation-value color space. Pixel-level segmentation masks remove background pixels before color analysis. The system was developed on a low-cost laptop that has only a central processing unit, and it was tested on still images, webcam input, and images captured by a DJI Mini 4K drone. Iterative testing exposed three classification faults: confusion between brown and orange at a hue boundary, a rule-ordering error that let the navy-blue rule capture black objects, and an under-constrained teal rule that labeled a white car as teal under cool lighting. Each fault was traced to a specific cause and corrected. After tuning, all 16 color classes were classified correctly in the final test images, including two overhead drone images of a person in a black outfit and a person in a white shirt. Because the same images were used for tuning and for reporting, this result demonstrates feasibility and is not an estimate of accuracy on unseen images. A held-out evaluation is identified as the next step.