GCS-YOLOV4-Tiny: A lightweight group convolution network for multi-stage fruit detection.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Traditional fruit growth stage assessment is manual, time-consuming, and labor-intensive.
• Visual assessment is hindered by fruit variations (size, color) and occlusions from leaves or branches.
• Automated detection of fruit growth stages is crucial for optimizing agricultural practices.

Purpose of the Study:

• To develop an efficient and accurate model for automatic detection of fruit growth stages.
• To enhance the YOLOV4-Tiny model for improved performance in fruit detection tasks.
• To reduce model size while maintaining or improving detection accuracy and speed.

Main Methods:

• Proposed the GCS-YOLOV4-Tiny model, integrating Squeeze and Excitation (SE) and Spatial Pyramid Pooling (SPP) modules.
• Utilized group convolution to reduce model size and increase detection speed.
• Evaluated the model on three public fruit datasets: Mango YOLO, Rpi-Tomato, and F. margarita.

Main Results:

• GCS-YOLOV4-Tiny demonstrated favorable performance (mAP, Recall, F1-Score, Average IoU) on Mango YOLO and Rpi-Tomato datasets.
• Achieved a model size of 20.70 MB with high accuracy metrics (mAP 93.42%, F1-score 90.80%) on the F. margarita dataset.
• Outperformed the state-of-the-art YOLOV4-Tiny model, showing a 17.45% increase in mAP and a 13.80% increase in F1-score.

Conclusions:

• The GCS-YOLOV4-Tiny model offers an effective and efficient solution for detecting fruit growth stages.
• The model's improvements in accuracy and speed make it suitable for real-world agricultural applications.
• The approach is extensible for detecting various fruits, crops, objects, or diseases.