Research on sugarcane stem counting method based on improved RT-DETR-seg

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate sugarcane counting is crucial for yield estimation, resource management, and harvesting optimization.
• Challenges in sugarcane fields include elongated plant shapes and complex occlusions, hindering precise counting.
• Existing methods struggle with feature fragmentation and environmental variations like lighting and shadows.

Purpose of the Study:

• To develop an improved instance segmentation model for accurate sugarcane counting in complex field conditions.
• To address challenges posed by sugarcane's elongated shape and occlusion for enhanced detection and segmentation.
• To provide a robust and efficient automated monitoring solution for slender economic crops.

Main Methods:

• An improved RT-DETR-seg (Real-Time Detection Transformer instance segmentation) model was developed.
• Incorporated asymmetric strip convolution (ASC) for enhanced vertical feature aggregation and noise suppression.
• Integrated coordinate attention (CA) mechanism to handle long-range dependencies and occluded regions.
• Introduced a joint moment-gradient regularized (MGR) loss function for improved segmentation accuracy under challenging lighting.

Main Results:

• The improved model achieved high accuracy: 91.03% for bounding box detection and 81.04% for mask segmentation.
• Demonstrated significant reduction in counting error, with Mean Absolute Error (MAE) decreasing from 1.09 to 0.28.
• Ablation studies confirmed the performance gains from ASC, CA mechanism, and MGR loss function individually.

Conclusions:

• The enhanced RT-DETR-seg model effectively addresses occlusion and feature fragmentation in sugarcane fields.
• The algorithm offers a balance of robust perception and real-time inference for automated crop monitoring.
• Provides a strong foundation for edge-computing systems for high-throughput, precision agriculture.