YOLOv9-LSBN: An Improved YOLOv9 Model for Cotton Pest and Disease

YOLOv9-LSBN: An Improved YOLOv9 Model for Cotton Pest and Disease

To achieve accurate identification of cotton aphids and diseases in natural complex environments, an enhanced YOLOv9 model named YOLOv9-LSBN (Large Selective Kernel Network with Bidirectional Feature Pyramid) is proposed. The RepLanLsk module replaces RepNCSPELAN4 in YOLOv9, dynamically adjusting re...

Full description

Saved in:
Bibliographic Details
Main Authors: Ruohong He, Fengkui Zhang, Jikui Zhu, Yulong Wang, Daorina Yang, Ting Zhang, Ping Li
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Cotton aphids
cotton disease
detection
RepLanLsk
BIFPN
YOLOv9
Online Access:https://ieeexplore.ieee.org/document/11031471/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:To achieve accurate identification of cotton aphids and diseases in natural complex environments, an enhanced YOLOv9 model named YOLOv9-LSBN (Large Selective Kernel Network with Bidirectional Feature Pyramid) is proposed. The RepLanLsk module replaces RepNCSPELAN4 in YOLOv9, dynamically adjusting receptive fields through reparameterizable convolutions and spatial attention, enhancing multi-scale feature diversity while maintaining computational efficiency via structural reparameterization during inference. A simplified weighted bidirectional feature pyramid network is integrated to strengthen bidirectional feature fusion and suppress background interference. Experiments on cotton aphids, Verticillium wilt, brown spot, and healthy leaves show that YOLOv9-LSBN achieves 93.1% precision, 92.5% recall, and 96.3% mAP@0.5, outperforming the original YOLOv9 by 1.8%, 0.3%, and 1.0%, respectively. Compared to YOLOv7, YOLOv8x, and lightweight models (e.g., YOLOv12), YOLOv9-LSBN demonstrates superior accuracy in complex backgrounds (96.3% vs. 94.6% mAP@0.5) with lower misjudgment rates, balancing real-time detection speed (28.3 ms/frame) and precision. Statistical significance (p <0.01) is validated through five randomized training trials, ensuring the robustness of improvements. This work provides a robust reference for crop pest detection in natural environments.
ISSN:2169-3536

Similar Items