Exploring Model Quantization in GenAI-based Image Inpainting and Detection of Arable Plants

2025-03-04

AI Summary

To address the dual challenges of scarce training data and limited computational resources on edge devices in agricultural weed detection, this paper proposes a lightweight co-optimization framework tailored for intelligent weeding systems. Methodologically: (1) we introduce a progressive image-inpainting data augmentation strategy based on Stable Diffusion, achieving up to +200% sample expansion in 10%-increment steps; (2) we pioneer a unified FP16/INT8 quantization pipeline jointly optimizing a generative AI inpainting model and YOLOv11/RT-DETR detectors. The framework is deployed end-to-end on Jetson Orin Nano, yielding substantial mAP₅₀ improvement, <1.2% accuracy degradation under INT8 quantization, and 2.3× inference speedup—enabling real-time operation. This work represents the first integration of generative data augmentation with multi-model collaborative quantization for field-level plant detection, establishing a reproducible technical paradigm for precision agriculture AI under resource-constrained conditions.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Search and Optimization: Learning to Search
• Computer Vision: Diffusion Models for Vision

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Web Mining and Content Analysis: Large pretrained models with web data
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI

Abstract

Deep learning-based weed control systems often suffer from limited training data diversity and constrained on-board computation, impacting their real-world performance. To overcome these challenges, we propose a framework that leverages Stable Diffusion-based inpainting to augment training data progressively in 10% increments -- up to an additional 200%, thus enhancing both the volume and diversity of samples. Our approach is evaluated on two state-of-the-art object detection models, YOLO11(l) and RT-DETR(l), using the mAP50 metric to assess detection performance. We explore quantization strategies (FP16 and INT8) for both the generative inpainting and detection models to strike a balance between inference speed and accuracy. Deployment of the downstream models on the Jetson Orin Nano demonstrates the practical viability of our framework in resource-constrained environments, ultimately improving detection accuracy and computational efficiency in intelligent weed management systems.

Problem

• Enhance training data diversity for weed control systems.
• Optimize model quantization for speed and accuracy.
• Deploy efficient models on resource-constrained devices.

Innovation

• Uses Stable Diffusion for data augmentation
• Applies FP16 and INT8 quantization strategies
• Deploys on Jetson Orin Nano for efficiency