Generative AI-based pipeline architecture for increasing training efficiency in intelligent weed control systems

In automated crop protection tasks, deep learning has demonstrated significant potential. However, these advanced models rely heavily on high-quality, diverse datasets, which are often scarce and costly to obtain in agricultural settings. Traditional data augmentation techniques, while useful for increasing the volume of the dataset, often fail to capture the real-world variability needed for robust model training. In this paper, we present a novel method for generating synthetic images to enhance the training of deep learning-based object detection models for intelligent weed control, aiming to improve data efficiency. The architecture of our GenAI-based image generation pipeline integrates the Segment Anything Model (SAM) for zero-shot domain adaptation with a text-to-image Stable Diffusion Model, enabling the creation of synthetic images that can accurately reflect the idiosyncratic properties and appearances of a variety of real-world conditions. We further assess the application of these synthetic datasets on edge devices by evaluating state-of-the-art lightweight YOLO models, measuring data efficiency by comparing mAP50 and mAP50-95 scores among different proportions of real and synthetic training data. Incorporating these synthetic datasets into the training process has been found to result in notable improvements in terms of data efficiency. For instance, most YOLO models that are trained on a dataset consisting of 10% synthetic images and 90% real-world images typically demonstrate superior scores on mAP50 and mAP50-95 metrics compared to those trained solely on real-world images. The integration of this approach opens opportunities for achieving continual self-improvement of perception modules in intelligent technical systems.