A Case for the Use of Chroma Cartesian Colour Representations for Image Classification on Plant-Based Domains

Alexis J S Payne, Gail Hopkins, Shreyank N Gowda, Isaac Triguero, Michael P Pound; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 7187-7196

The primary approach in computer vision across all domains is to work with the Red, Green, Blue (RGB) colour representation. This is most clear in the use of transfer learning, where leveraging known good weights for a model, pretrained on large datasets in RGB, is commonplace. Despite the dominance of RGB, alternative colour representations, such as Hue, Saturation, Value (HSV), have seen use in plant-based domain specific tasks, or with models that are trained from the ground up. In this paper we have explored a wide array of known colour representations, across a variety of plant and plant disease identification datasets. These show that there is not just several colour representations that consistently outperform RGB in training from scratch, but a common construction of the representations that do: an isolated Luminosity channel, similar to a black and white version of the image;combined with a Cartesian representation of the Chroma component, the colour information. We propose a new and effective addition to this class, H2SV, which derives from the relatively widely used HSV representation, improving upon its polar representation of colour information and limited effectiveness when used with neural networks. We conduct extensive experiments that show the effectiveness of these colour representations over RGB is not limited to a single architecture, and this trend is seen across almost all plant datasets. We also show that these benefits can be combined with appropriately pretrained weights so both the computer vision norm of transfer learning, and these alternative colour representations can be leveraged at once. Relevant code tooling is available via PyPI.