Automated Generation of Accurate, Compact and Focused Crop and Weed Segmentation Models

Abstract

In this work, we address a challenging problem in crop sciences, where on-device execution of machine inference is necessary. There are well-crafted large machine learning models for crop and weed segmentation. However, the lack of focus on building compact yet accurate models is deterring the success of precision agriculture. Designing such models not only needs specialized cross-disciplinary skills in machine learning and embedded systems but also requires tedious and repetitive trials. To address this problem, we propose a method to automatically tune and generate accurate and optimal models for Precision Agriculture. These models are not only suitable for on-device inference, but also enhance the performance of selected crops (or weeds) in a multi-class segmentation setting. Extensive experimental evaluations are performed to demonstrate the efficacy of the proposed search on more than one crop and weed segmentation dataset. The optimized models perform better than the state-of-the-art models used in crop and weed segmentation, using only a fraction of the parameters. We believe that the fast and automated generation of optimal neural models has the potential to address the requirement of on-device inference in precision agriculture.