PAM-FOG Net: A Lightweight Weed Detection Model Deployed on Smart Weeding Robots

Abstract

Visual target detection based on deep learning with high computing power devices has been successful, but the performance in intelligent agriculture with edge devices has not been prominent. Specifically, the existing model architecture and optimization methods are not well-suited to low-power edge devices, the agricultural tasks such as weed detection require high accuracy, short inference latency, and low cost. Although there are automated tuning methods available, the search space is extremely large, using existing models for compression and optimization greatly wastes tuning resources. In this article, we propose a lightweight PAM-FOG net based on weed distribution and projection mapping. More significantly, we propose a novel model compression optimization method to fit our model. Compared with other models, PAM-FOG net runs on smart weeding robots supported by edge devices, and achieves superior accuracy and high frame rate. We effectively balance model size, performance and inference speed, reducing the original model size by nearly 50%, power consumption by 26%, and improving the frame rate by 40%. It shows the effectiveness of our model architecture and optimization method, which provides a reference for the future development of deep learning in intelligent agriculture.