Exploring Model Quantization in GenAI-based Image Inpainting and Detection of Arable Plants
This work addresses computational efficiency and data scarcity for intelligent weed management systems, representing an incremental improvement by combining existing methods like inpainting and quantization in a new application.
The paper tackled limited training data and computational constraints in deep learning-based weed control by proposing a framework that uses Stable Diffusion-based inpainting to augment training data by up to 200%, evaluated on YOLO11(l) and RT-DETR(l) detection models with mAP50 metrics, and explored quantization (FP16 and INT8) to improve inference speed and accuracy on Jetson Orin Nano.
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.