S-YOLOv10-SIC

Citation Author(s):
Wang
Baijuan
Zhang
Shihao
Submitted by:
Wang Baijuan
Last updated:
Thu, 09/12/2024 - 12:12
DOI:
10.21227/48v5-6x60
License:
47 Views
Categories:
Keywords:
0
0 ratings - Please login to submit your rating.

Abstract 

To solve the problem of accurate recognition and picking of tea by tea picking robot, this study proposes a S-YOLOv10-SIC algorithm that integrates slice-assisted hyper-inference algorithm. This algorithm enhances the YOLOv10 network by introducing Space-to-Depth Convolution, asymptotic feature pyramid network, and Inner-IoU. These improvements reduce the loss of detailed information in long-distance and low-resolution images, improve key layer saliency, optimize non-adjacent layer fusion, enhance model convergence speed, and increase model universality. Experimental results demonstrate significant enhancements compared to YOLOv10, including a reduction of more than 30% in Bounding Box Regression Loss in the training set and a reduction of more than 60% in Classification Loss and Bounding Box Regression Loss in the verification set. Precision, Recall, and mAP also increased by 7.1%, 6.69%, and 6.78% respectively. Additionally, AP values for specific tea types saw improvements, with single bud, one bud and one leaf, and one bud and two leaves increasing by 6.10%, 7.99%, and 8.28% respectively. These advancements allow the improved model to effectively handle long-distance, small target, and low-resolution challenges while maintaining high precision and recall rates, laying the groundwork for the development of an Anji white tea picking robot.

Instructions: 

To solve the problem of accurate recognition and picking of tea by tea picking robot, this study proposes a S-YOLOv10-SIC algorithm that integrates slice-assisted hyper-inference algorithm. This algorithm enhances the YOLOv10 network by introducing Space-to-Depth Convolution, asymptotic feature pyramid network, and Inner-IoU. These improvements reduce the loss of detailed information in long-distance and low-resolution images, improve key layer saliency, optimize non-adjacent layer fusion, enhance model convergence speed, and increase model universality. Experimental results demonstrate significant enhancements compared to YOLOv10, including a reduction of more than 30% in Bounding Box Regression Loss in the training set and a reduction of more than 60% in Classification Loss and Bounding Box Regression Loss in the verification set. Precision, Recall, and mAP also increased by 7.1%, 6.69%, and 6.78% respectively. Additionally, AP values for specific tea types saw improvements, with single bud, one bud and one leaf, and one bud and two leaves increasing by 6.10%, 7.99%, and 8.28% respectively. These advancements allow the improved model to effectively handle long-distance, small target, and low-resolution challenges while maintaining high precision and recall rates, laying the groundwork for the development of an Anji white tea picking robot.

Comments

ok

Submitted by Chala Abdissa on Fri, 09/13/2024 - 18:23