Improving wheat head detection: a data-centric approach by domain variance reduction

Abstract

AI-assisted agriculture helps in obtaining accurate data about a worldwide staple food - wheat. Recently computer vision has been employed to detect wheat heads. It is a challenging task as wheat heads are small and vary in appearance due to on their growth phase, field conditions and planting patterns. This paper presents a data-driven approach to reduce variance among wheat head images to improve detection results. We use preprocessing steps to standardize the color histogram of images and train a YOLOv5 model to detect the wheat heads. We use histogram matching to make a color histogram of every image consistent with the global average color histogram of the training set. This is followed by contrast-limited adaptive histogram equalization (CLAHE) to improve the contrast on a local scale and improve the visibility of the wheat heads. With 0.715 Average Domain Accuracy (ADA), our approach achieves state-of-the-art (SOTA) results, outperforming the existing SOTA (0.700 ADA) approach by 1.5% while requiring almost 5x fewer training computation steps.