A System for Weeds and Crops Identification-Reaching over 10 FPS on Raspberry Pi with the Usage of MobileNets,...

Area of Science:

• Agrorobotics
• Computer Vision
• Deep Learning

Background:

• Automated weeding is crucial in agriculture for mechanical or herbicide-based weed removal.
• Deploying deep learning for computer vision tasks on low-cost mobile computers remains a challenge.

Purpose of the Study:

• To develop an efficient deep learning system for automated weed detection and segmentation on low-cost hardware.
• To improve the accuracy and speed of weed identification in agricultural settings.

Main Methods:

• A custom Convolutional Neural Network (CNN) architecture combining U-Net, MobileNets, DenseNet, and ResNet concepts was developed.
• Knowledge distillation was used to reduce the need for manual ground truth labels.
• Custom modifications, including separable convolutions and reduced channel numbers, were implemented to decrease inference time.

Main Results:

• The system achieved satisfying accuracy, detecting 47-67% of weed area while misclassifying 0.1-0.9% of crop area.
• The system operated at over 10 frames per second on a Raspberry Pi 3B+.
• The system was tested on four plant species under various growth stages and lighting conditions.

Conclusions:

• The developed system offers an efficient solution for automated weeding using deep learning on resource-constrained devices.
• The novel techniques for model optimization and data labeling can be applied to other agrorobotics applications.