Abstract
Abstract
Fruit detection and segmentation are key components of agricultural automation pipelines, enabling tasks such as yield estimation, quality sorting, and robotic harvesting in resource-constrained environments. This study evaluates four colour spaces – RGB, HSV, CIE L*a*b* (LAB), and YCbCr – for threshold-based fruit segmentation in precision agriculture and robotic harvesting. It specifically investigates which colour space maintains the highest accuracy and stability under variable indoor illumination. A dataset of 144 images featuring fruits from four colour categories (red, green, orange, and yellow) was utilised. Colour thresholds were calibrated through manual region-of-interest selection and applied using chromatic thresholding and morphological refinement, resulting in 2304 mask evaluations based on intersection over union (IoU), precision, recall, and F1-score. Clear and statistically significant differences were found among colour spaces: HSV achieved the highest overall performance (F1-score = 0.972, IoU = 0.947), followed by RGB (F1-score = 0.823), LAB (F1-score = 0.714), and YCbCr (F1-score = 0.682), confirmed by non-parametric statistical testing. HSV is established as the most accurate and consistent colour representation for fruit segmentation under the proposed experimental protocol, providing practical guidance for lightweight, real-time segmentation pipelines in agricultural automation systems operating under variable illumination conditions.