An improved YOLOv7 model based on Swin Transformer and Trident Pyramid Networks for accurate tomato detection

Area of Science:

• Computer Vision and Machine Learning
• Agricultural Robotics
• Image Processing

Background:

• Accurate fruit detection is essential for automated fruit picking, but environmental challenges like varying illumination and occlusion hinder performance.
• Existing fruit detection models struggle with real-world complexities, impacting the viability of fruit harvesting robots.
• The commercialization of agricultural robotics is directly linked to advancements in detection accuracy under diverse conditions.

Purpose of the Study:

• To develop an improved fruit detection model addressing challenges in complex environmental conditions.
• To enhance the accuracy and robustness of automated tomato detection for robotic harvesting applications.
• To introduce novel architectural components and loss functions for superior object detection performance.

Main Methods:

• Proposed YOLO-SwinTF model, an adaptation of YOLOv7, incorporating Swin Transformer (ST) blocks for enhanced global information capture.
• Integration of Trident Pyramid Networks (TPN) to improve feature map communication and processing beyond standard PANet.
• Introduction of Focaler-IoU loss function to dynamically adjust focus based on sample difficulty, optimizing training.

Main Results:

• The YOLO-SwinTF model achieved high detection metrics on a tomato dataset: 96.27% recall, 96.17% precision, 96.22% F1-score, and 98.67% AP.
• Demonstrated significant performance improvements over the original YOLOv7 model, with notable gains in recall, precision, F1-score, and AP.
• Exhibited superior accuracy compared to other state-of-the-art methods while maintaining competitive detection speed and strong robustness to lighting and occlusion.

Conclusions:

• The YOLO-SwinTF model effectively addresses the limitations of current fruit detection systems in complex agricultural environments.
• The integration of Swin Transformer blocks and TPN significantly boosts the model's ability to handle long-range dependencies and feature communication.
• The proposed model shows substantial potential for practical application in automated tomato harvesting, offering high accuracy and robustness.