ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

A Diffusion Model-Based Restoration Framework for Robust Visual Perception in Agricultural Systems

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600932.(doi:10.13031/aim.202600932)
Authors:   Linyuan Wang, Haibo Yao, Yanbo Huang, Te-Ming Tseng, Xin Sun, Dong Chen
Keywords:   Image restoration, diffusion models, generative models, precision agriculture, super-resolution, weed detection

Abstract. Recent advances in robotics, artificial intelligence, and sensing technologies have significantly accelerated the development of intelligent agricultural systems. Central to these systems are machine vision components, which enable automated perception by extracting and interpreting visual information from field imagery for tasks such as detection, classification, and decision-making. However, the performance of these vision systems in real-world agricultural environments is often compromised by image degradation caused by hardware limitations and environmental factors, such as motion blur, noise, and lighting. These degradations lead to the loss of fine-grained visual details, thereby reducing the reliability of downstream perception tasks. To address this challenge, we propose an integrated framework that leverages diffusion-based image restoration to enhance degraded field imagery for robust visual recognition. Specifically, the framework first employs Diffusion-Based Image Restoration (DiffBIR) to reconstruct high-quality images from degraded inputs, followed by a downstream detection model to perform task-specific perception. We validate the proposed framework using multi-class weed detection as a representative use case with the YOLOv10 detector. Experimental results demonstrate that the proposed approach significantly improves detection accuracy and maintains botanical details under degraded conditions, outperforming traditional interpolation (Bicubic) and state-of-the-art deep learning-based super-resolution methods, including GAN-based and Transformer-based methods. Furthermore, when YOLOv10 was applied to DiffBIR-enhanced images, detection performance improved by over 50% in both mAP@0.5 and mAP@0.5:0.95 across all evaluated model scales. These results highlight the potential of diffusion-based restoration as a general-purpose enhancement module for improving the robustness and reliability of machine vision systems under real-world degradation conditions.

(Download PDF)    (Export to EndNotes)