Fig-YOLO: An Improved YOLOv11-Based Fig Detection Algorithm for Complex Environments.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate fig detection is crucial for intelligent harvesting but hindered by small targets, occlusion, and similar backgrounds.
• Existing methods struggle with the complexities of real-world orchard environments.

Purpose of the Study:

• To develop an improved YOLOv11n-based algorithm, named Fig-YOLO, for enhanced fig detection.
• To address key challenges including small object size, occlusion, and background similarity.

Main Methods:

• Introduced Spatial-Frequency Selective Convolution (SFSConv) in the backbone for joint spatial and frequency feature modeling.
• Incorporated an enhanced bi-branch attention mechanism (EBAM) to improve representation of key regions under occlusion.
• Utilized a multi-branch dynamic sampling convolution (MFCV) module for capturing figs of varying sizes and feature fusion.

Main Results:

• Fig-YOLO achieved high performance metrics: 89.2% precision, 78.4% recall, and 87.3% mAP@0.5.
• Demonstrated substantial improvement over the baseline YOLOv11n algorithm.
• Maintained stable performance across diverse conditions (fruit size, occlusion, lighting, data sources).

Conclusions:

• Fig-YOLO effectively overcomes major obstacles in fig detection for intelligent harvesting.
• The proposed architectural innovations enhance robustness and accuracy in complex environments.
• This algorithm provides strong support for automated orchard monitoring and harvesting systems.