Dissimilarity-based active learning for embedded weed identification

Abstract

Weed identification helps ensure crop yield and realize precision agriculture. Although the deep learning-based methods have achieved high performance, their needed large-scale annotated data is difficult to obtain, and the massive parameters lead to difficulties in model deployment in embedded applications. To develop efficient crop weeds classification system, we propose a dissimilarity-based method to select few but representative samples and consider data diversity. Many experiments are conducted on two different datasets. Results show that the proposed method can obtain test accuracy of 90.75% with 32% amount of the DeepWeeds dataset, up to 99.18% of the baseline accuracy. For the Grass-Broadleaf dataset, our approach can obtain the test accuracy of 98.97% with 27.8% amount of the dataset, up to 99.5% of the baseline accuracy. Further, the designed model is compressed from 117.9 MB to 8.6 MB, with a compression ratio of 92.7%, while with only 1% degradation of accuracy. Finally, the compressed model is deployed to NVIDIA Jetson AGX Xavier, running at 192 fps. This work can lay a foundation for the following research about few-shot learning, data quality assessment, and model deployment in agriculture.