Exploiting image entropy and distance normalization for fruit sizing

Fruit size and its evolution during the growing season are key parameters for effective orchard management and the economic viability in horticulture. Estimating fruit size and growth involves measuring fruit dimensions and weight over time. Traditional methods, such as manual sampling with calipers, and advanced plant-based sensors, are limited in their application by labour intensity, high costs and reduced sample size. Computer vision (CV) algorithms offer a promising alternative, enabling the monitoring of large areas with increased reliability and sample size. Mounted on scanning platforms or tractors, these systems leverage recent advances in CV algorithms to detect and measure fruit before, during, and after harvest. This study evaluates a CV system using a consumer-grade depth camera (RGB-D) for data collection, manual annotations for fruit detect and the proposed algorithm for fruit size estimation. The proposed approach combines distance normalization and image lightness local entropy to enhance size estimation. Results indicate a mean error of 0.8 mm, and an RMSE of ~4.0 mm, with a correlation coefficient of r=0.96 in fruit. This study demonstrated improved accuracy over previous methods relying solely on colour features for the shape fitting. This innovative CVS approach highlights the potential for cost-effective, scalable, and precise fruit monitoring solutions.