Research on Field Weed Target Detection Algorithm Based on Deep Learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning weed detection is vital for smart agriculture.
• YOLO algorithms offer efficiency but struggle with accuracy for occluded weeds.
• Existing methods lack robustness in complex agricultural scenarios.

Purpose of the Study:

• To develop an advanced weed detection algorithm for improved accuracy in smart agriculture.
• To address the limitations of current YOLO algorithms in detecting occluded or overlapping weeds.
• To enhance the efficiency and robustness of weed detection systems.

Main Methods:

• Proposed SSS-YOLO algorithm based on YOLOv9t.
• Introduced Spatial Channel Conv Block (SCB) for long-range dependency capture and feature enhancement.
• Developed Spatial Pyramid Pooling Fast Edge Gaussian Aggregation Super (SPPF EGAS) for multi-scale feature extraction and background inference.
• Implemented Efficient Multi-Scale Spatial-Feedforward Network (EMSN) for semantic reconstruction and interference suppression.

Main Results:

• SSS-YOLO demonstrated significant performance improvements over existing algorithms.
• The proposed SCB, SPPF EGAS, and EMSN modules effectively handled weed occlusion and overlap.
• Experiments on custom and public datasets (Cotton WeedDet12) validated the method's efficacy.

Conclusions:

• SSS-YOLO offers a robust solution for accurate weed detection in challenging agricultural environments.
• The novel modules enhance the capability of deep learning models for smart farming applications.
• This research contributes to advancing precision agriculture through improved automated weed management.