LGN-YOLO: A Leaf-Oriented Region-of-Interest Generation Method for Cotton Top Buds in Fields
As small-sized targets, cotton top buds pose challenges for traditional full-image search methods, leading to high sparsity in the feature matrix and resulting in problems such as slow detection speeds and wasted computational resources. Therefore, it is difficult to meet the dual requirements of re...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-06-01
|
| Series: | Agriculture |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2077-0472/15/12/1254 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | As small-sized targets, cotton top buds pose challenges for traditional full-image search methods, leading to high sparsity in the feature matrix and resulting in problems such as slow detection speeds and wasted computational resources. Therefore, it is difficult to meet the dual requirements of real-time performance and accuracy for field automatic topping operations. To address the low feature density and redundant information in traditional full-image search methods for small cotton top buds, this study proposes LGN-YOLO, a leaf-morphology-based region-of-interest (ROI) generation network based on an improved version of YOLOv11n. The network leverages young-leaf features around top buds to determine their approximate distribution area and integrates linear programming in the detection head to model the spatial relationship between young leaves and top buds. Experiments show that it achieves a detection accuracy of over 90% for young cotton leaves in the field and can accurately identify the morphology of young leaves. The ROI generation accuracy reaches 63.7%, and the search range compression ratio exceeds 90%, suggesting that the model possesses a strong capability to integrate target features and that the output ROI retains relatively complete top-bud feature information. The ROI generation speed reaches 138.2 frames per second, meeting the real-time requirements of automated topping equipment. Using the ROI output by this method as the detection region can address the problem of feature sparsity in small targets during traditional detection, achieve pre-detection region optimization, and thus reduce the cost of mining detailed features. |
|---|---|
| ISSN: | 2077-0472 |