Agricultural weed identification using optimized deep learning architecture

Rai, Nitin; Zhang, Yu; Villamil, Maria; Howatt, Kirk; Ostlie, Michael; Sun, Xin

Bckground

The work presented in this research uses state-of-the-art YOLOv7 tiny architecture and integrates two optimization techniques within the backbone and neck components. These two optimization techniques are: (a) module re-parameterized convolutional layer (RCL), and (b) filter-based structured pruning (model compression). Both of these techniques have been adopted from here and here. For more details read the paper here and browse through the repository to download the metric files and other assessment docs. The developed architecture uses 78% less parameters compared to the YOLOv7-Base model and has been built and optimized for weed detection and localization for site-specific weed management application.

Novel aspects of this research study are:

1. Integration of “RepConv” (Re-parameterized Convolutional Module in the neck compoenent)
2. Filter-based structured pruning approach
3. Conversion to FP16 (floating points) for edge deployment

Highlights of this research work includes:

1. Open-source dataset: 3,929 images and 12k bounding-box annotations used in this study.
2. Deep learning model optimization using pruning and integrating re-parameterization module.
3. Assessing the effects of training multiple image resolution on the optimized YOLO-Spot model.
4. YOLO-Spot_M model achieves accuracy (+1.3%) and mAP (+2.7%) compared to YOLO-base model.
5. YOLO-Spot_M with half-precision gains 5X times inferencing speed on an edge device.

Hardware used to perform training and inference tasks:

Reference

Rai, N., Zhang, Y., Villamil, M., Howatt, K., Ostlie, M., & Sun, X. (2024). Agricultural weed identification in images and videos by integrating optimized deep learning architecture on an edge computing technology . Comput. Electron. Agric., 216, 108442. https://doi.org/10.1016/j.compag.2023.108442