An efficient tomato-detection method based on improved YOLOv4-tiny model in complex environment.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Greenhouse fruit detection is hindered by environmental factors like occlusion and varying illumination.
• Existing methods struggle with accuracy due to complex conditions such as overlapping and clustered fruits.

Purpose of the Study:

• To develop an accurate and robust fruit detection algorithm for tomatoes in greenhouses.
• To improve the efficiency and accuracy of the YOLOv4-tiny model for real-time applications.

Main Methods:

• An improved YOLOv4-tiny model was proposed, featuring a modified backbone network replacing BottleneckCSP modules with Bottleneck and reduced BottleneckCSP modules.
• A tiny version of CSP-Spatial Pyramid Pooling (CSP-SPP) was integrated to enhance the receptive field.
• The Content Aware Reassembly of Features (CARAFE) module replaced traditional up-sampling in the neck for improved high-resolution feature maps.

Main Results:

• The improved YOLOv4-tiny model achieved high performance metrics: 96.3% precision, 95% recall, 95.6% F1-score, and 82.8% mean average precision (mAP) at IoU 0.5-0.95.
• The model demonstrated a rapid detection time of 1.9 ms per image.
• Performance surpassed state-of-the-art methods, meeting real-time tomato detection requirements.

Conclusions:

• The enhanced YOLOv4-tiny model offers superior efficiency and accuracy for greenhouse tomato detection.
• The proposed modifications effectively address environmental challenges, enabling reliable real-time fruit identification.