BFP Net: Balanced Feature Pyramid Network for Small Apple Detection in Complex Orchard Environment.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Small fruit detection, particularly for immature green fruits, presents significant challenges in natural orchard environments.
• The visual similarity between small fruit skin and background complicates accurate localization.
• Existing methods struggle with the low pixel count and subtle features of small targets.

Purpose of the Study:

• To propose a novel deep learning model, the Balanced Feature Pyramid Network (BFP Net), for enhanced small apple detection.
• To address the information imbalance and spatial localization issues inherent in detecting small fruits.
• To improve the accuracy and generalization performance of automated fruit detection systems.

Main Methods:

• Developed a Balanced Feature Pyramid Network (BFP Net) integrating multi-scale features from FPN layers.
• Introduced a novel extended feature layer from ResNet50 conv1 and a decoupled-aggregated module to enhance spatial information.
• Implemented a weight-like feature fusion architecture and Kullback-Leibler distillation loss for knowledge transfer.

Main Results:

• Achieved Average Precision (AP_S) scores of 47.0% on GreenApple, 42.2% on MinneApple, and 35.6% on Pascal VOC datasets.
• Demonstrated superior performance compared to several state-of-the-art small fruit detection methods.
• Validated the model's robustness and good generalization capabilities across different datasets.

Conclusions:

• The proposed BFP Net effectively tackles the challenges of small fruit detection by balancing multi-scale information and enhancing spatial localization.
• The novel architectural components and training strategy contribute to significant improvements in detecting small, immature apples.
• This research offers a promising advancement for precision agriculture and automated harvesting systems.