Grape clusters detection based on multi-scale feature fusion and augmentation.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Grape cluster detection faces challenges due to scale variations, illumination changes, and occlusion in real-world agricultural settings.
• Existing methods often struggle with accuracy in complex and variable environments, impacting yield estimation and management.

Purpose of the Study:

• To improve the accuracy and robustness of grape cluster detection in challenging agricultural scenes.
• To develop an enhanced YOLOv7 network capable of handling scale differences, illumination variations, and object occlusion.

Main Methods:

• Proposed a novel YOLOv7 network incorporating a Multi-Scale Feature Extraction Module (MSFEM) for small targets.
• Introduced a Receptive Field Augmentation Module (RFAM) using dilated convolution to improve detection across various scales.
• Integrated a Spatial Pyramid Pooling Cross Stage Partial Concatenation Faster (SPPCSPCF) module for multi-scale feature fusion and faster training.
• Incorporated a Residual Global Attention Mechanism (ResGAM) to focus on critical features and regions.

Main Results:

• Achieved a mean Average Precision (mAP) of 93.29% on the GrappoliV2 dataset, a 5.39% improvement over standard YOLOv7.
• Increased Precision by 2.83%, Recall by 3.49%, and F1 score by 0.07 compared to the baseline YOLOv7 model.
• Demonstrated superior detection performance and adaptability compared to state-of-the-art methods in varied environmental conditions.

Conclusions:

• The proposed multi-scale feature fusion and augmentation YOLOv7 network significantly enhances grape cluster detection accuracy.
• The method effectively addresses challenges like scale variation and occlusion, proving its adaptability for agricultural computer vision.
• This advanced detection system offers a promising solution for precision agriculture, improving monitoring and management of grape cultivation.