Apple recognition in complex environments based on FC-DETR.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Apple cultivation in Xinjiang faces labor shortages, driving demand for automated harvesting.
• Current Convolutional Neural Network (CNN) models struggle with complex environments due to limited global feature capture.
• Accurate and robust apple recognition is critical for practical automated harvesting solutions.

Purpose of the Study:

• To develop a real-time object detection model for accurate apple recognition in complex environments.
• To overcome the limitations of existing CNN models in capturing global features.
• To enhance the efficiency and sustainability of the apple industry through advanced technology.

Main Methods:

• Proposed the FC-DETR (Feature-enhanced Contextual Detection Transformer) model.
• Incorporated the FEMA-BasicBlock residual module for enhanced feature processing.
• Integrated the CAFM cross-scale adaptive feature fusion module for multi-scale feature selection.
• Utilized the Inner-WIoU loss function to improve detection accuracy.

Main Results:

• The FC-DETR model achieved an 87% apple recognition accuracy in complex backgrounds.
• A recall rate of 82% was obtained, demonstrating effective detection.
• The model maintained a lightweight design suitable for practical applications.

Conclusions:

• The FC-DETR model offers significant advancements in automated apple harvesting technology.
• This research contributes to improving efficiency and sustainability in the apple industry.
• The proposed model demonstrates high accuracy and robustness in challenging environments.