SemanticSugarBeets: A Multi-Task Framework and Dataset for Inspecting Harvest and Storage Characteristics of Sugar Beets

TL;DR

This work addresses automated quality assessment of post-harvest and post-storage sugar beets to preserve sugar content. It introduces SemanticSugarBeets, a large, multi-stage dataset with semantic and instance annotations across Sample, Harvest, and Storage stages, plus two-stage training: fast coarse beet detection and marker detection followed by fine-grained semantic segmentation inside beet interiors for detailed damage, soil, and vegetation labeling, along with mass estimation using scale markers. Extensive experiments evaluate multiple architectures, encoders, and image sizes under varying environmental conditions, reporting high detection performance (mAP50-95 ≈ 98.8) and competitive segmentation accuracy (mean IoU ≈ 64.0) and demonstrate the practicality of online quality control in production lines. The dataset and framework enable robust beet inspection, informing storage strategies and potential extensions to other crops, with future work focusing on broader conditions and more damage classes.

Abstract

While sugar beets are stored prior to processing, they lose sugar due to factors such as microorganisms present in adherent soil and excess vegetation. Their automated visual inspection promises to aide in quality assurance and thereby increase efficiency throughout the processing chain of sugar production. In this work, we present a novel high-quality annotated dataset and two-stage method for the detection, semantic segmentation and mass estimation of post-harvest and post-storage sugar beets in monocular RGB images. We conduct extensive ablation experiments for the detection of sugar beets and their fine-grained semantic segmentation regarding damages, rot, soil adhesion and excess vegetation. For these tasks, we evaluate multiple image sizes, model architectures and encoders, as well as the influence of environmental conditions. Our experiments show an mAP50-95 of 98.8 for sugar-beet detection and an mIoU of 64.0 for the best-performing segmentation model.

Figures (11)

• Figure 1: Representative images and semantic segmentation masks of the SemanticSugarBeets dataset from three processing stages. Label categories are listed in the color bar. The background class is displayed in black.
• Figure 2: Representative samples of annotated lighting and soil-moisture conditions as well as their distributions of beets across the three processing stages.
• Figure 3: Distributions of annotated pixels per processing stage for Beet, Cut, Leaf, Soil, Damage and Rot classes.
• Figure 4: Overview of our two-stage approach to detection and segmentation of sugar beets. First, beets are isolated through instance segmentation. Second, detected beet patches are finely segmented. Finally, segmentation results are fused to provide semantic segmentation of vegetation, soil adhesion, damage and cutting surfaces. Additionally, oriented bounding boxes of reference markers are detected and used for inferring the mass of each beet.
• Figure 5: Maximum semantic-segmentation performance for each architecture, encoder and input image size.