Improved RT-DETR and its application to fruit ripeness detection.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Manual crop maturity detection is inefficient and costly, hindering automated harvesting.
• Accurate crop maturity status recognition is crucial for optimizing agricultural practices and yield.

Purpose of the Study:

• To enhance the Real-Time DEtection TRansformer (RT-DETR) model for improved crop maturity detection.
• To increase detection accuracy while reducing model size and computational complexity.

Main Methods:

• Refined the Backbone structure of RT-DETR using HG Block Enhancement with Rep Block and Partial Convolution (PConv).
• Incorporated Efficient Multi-Scale Attention (EMA) to optimize feature distribution and model efficiency.
• Evaluated the enhanced model against the original RT-DETR and YOLOv8 on relevant datasets.

Main Results:

• The enhanced RT-DETR model demonstrated a 2.9% increase in average accuracy (mAP@0.5).
• Achieved a 5.5% reduction in model size and a 9.6% decrease in computational complexity.
• Outperformed existing models like YOLOv8 in general detection scenarios.

Conclusions:

• The proposed enhancements significantly improve RT-DETR's performance for crop maturity detection.
• The refined model offers a more accurate, efficient, and computationally less intensive solution for automated agriculture.
• This advancement supports the development of more effective automated harvesting systems.