Deep Learning-Based Image Analysis for Automated Fruit Sizing

Abstract

Accurate fruit sizing is a critical task in the agricultural supply chain, but current methods are either labor-intensive or rely on expensive, calibration-sensitive hardware. This study develops and evaluates a novel, low-cost deep learning framework for automated tomato circumference estimation using multiview RGB imagery and an embedded physical scale reference. A dataset of tomato images from two orthogonal viewpoints was captured using a smartphone, and a custom-designed gridded marker was applied to provide an embedded reference for scale and local curvature. A dual-branch convolutional neural network (CNN) based on EfficientNet-B1 was trained to regress circumference by fusing image features with extracted geometric metadata, such as curvature, eccentricity, and scale factor. The proposed dual-view model, which integrates the reference marker and extracted geometric features, achieved high accuracy, yielding an mean absolute error (MAE) of 3.70 mm (1.46% MAPE). This result represents a 55% improvement in MAE over the single-view baseline (8.19 mm MAE), and establishes a highly competitive relative percentage error (1.46% MAPE) that outperforms recent state of the art depth-sensing and artificial intelligence pipelines. The fusion of image features with metadata was found to be critical, as dual-view models without these features failed to converge effectively. The study concludes that the fusion of multiview geometric constraints with explicit, marker-derived physical scale cues is essential for achieving high-accuracy, low-cost fruit sizing. This framework presents a viable, scalable, and low-cost alternative to complex depth-sensing or stereo-vision systems, enabling field-deployable, millimeter-scale fruit sizing using only standard RGB cameras. Using this method, we achieve a 36% increase in accuracy compared to approaches relying on 3-D vision systems and outperform existing automated RGB vision systems.