YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection
Abstract Accurate monitoring of cotton maturity is crucial for improving both yield and quality in cotton production. However, existing models often suffer from low detection accuracy and poor adaptability due to challenges such as varietal diversity, rapid changes in maturity stages, difficulty in...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-07-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00081-8 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1849235681297039360 |
|---|---|
| author | Meiqi Zhong Linjing Wei Henghui Mo |
| author_facet | Meiqi Zhong Linjing Wei Henghui Mo |
| author_sort | Meiqi Zhong |
| collection | DOAJ |
| description | Abstract Accurate monitoring of cotton maturity is crucial for improving both yield and quality in cotton production. However, existing models often suffer from low detection accuracy and poor adaptability due to challenges such as varietal diversity, rapid changes in maturity stages, difficulty in small object detection, and environmental interferences like lighting and occlusion. To address these issues, this paper proposes a lightweight cotton maturity detection model named YOLOv8n-DSDL (Dual-backbone & Decoupled Focused Self-Attention & DySample & Lightweight) based on the YOLOv8 framework. The model integrates the original YOLOv8n with a lightweight StarNet to form a dual-backbone structure, enabling efficient multi-channel feature fusion and enhancing fine-grained maturity feature extraction. A novel Decoupled Focused Self-Attention (DFSA) mechanism is designed, which utilizes one-dimensional dilated convolutions and key-value fusion operations to dynamically enhance feature associations and improve perception of subtle visual cues such as color variation and micro-cracks. On the data side, a Mosaic9 augmentation strategy is employed in conjunction with StyleGAN3 and Laplacian variance-based filtering to generate high-quality synthetic images, thereby increasing dataset diversity. To mitigate image distortion and edge aliasing during upsampling, a dynamic upsampling operator, DySample, is introduced. After deployment on the Jetson Xavier NX platform with TensorRT acceleration, the model achieves a mean precision of 89.24%, a mean recall of 87.71%, and a mAP@50 of 87.02%, running at 43 FPS. The overall performance surpasses the original YOLOv8n and other mainstream detection methods. This study provides a novel and practical solution for intelligent maturity monitoring of cotton and potentially other crops. The source code is publicly available at: https://github.com/mohenghui/Cotton_maturity . |
| format | Article |
| id | doaj-art-6b28810741344743864f63d04bcd848a |
| institution | Kabale University |
| issn | 1319-1578 2213-1248 |
| language | English |
| publishDate | 2025-07-01 |
| publisher | Springer |
| record_format | Article |
| series | Journal of King Saud University: Computer and Information Sciences |
| spelling | doaj-art-6b28810741344743864f63d04bcd848a2025-08-20T04:02:42ZengSpringerJournal of King Saud University: Computer and Information Sciences1319-15782213-12482025-07-0137513010.1007/s44443-025-00081-8YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detectionMeiqi Zhong0Linjing Wei1Henghui Mo2College of Information Science and Technology, Gansu Agricultural UniversityCollege of Information Science and Technology, Gansu Agricultural UniversityCollege of Information Science and Technology, Gansu Agricultural UniversityAbstract Accurate monitoring of cotton maturity is crucial for improving both yield and quality in cotton production. However, existing models often suffer from low detection accuracy and poor adaptability due to challenges such as varietal diversity, rapid changes in maturity stages, difficulty in small object detection, and environmental interferences like lighting and occlusion. To address these issues, this paper proposes a lightweight cotton maturity detection model named YOLOv8n-DSDL (Dual-backbone & Decoupled Focused Self-Attention & DySample & Lightweight) based on the YOLOv8 framework. The model integrates the original YOLOv8n with a lightweight StarNet to form a dual-backbone structure, enabling efficient multi-channel feature fusion and enhancing fine-grained maturity feature extraction. A novel Decoupled Focused Self-Attention (DFSA) mechanism is designed, which utilizes one-dimensional dilated convolutions and key-value fusion operations to dynamically enhance feature associations and improve perception of subtle visual cues such as color variation and micro-cracks. On the data side, a Mosaic9 augmentation strategy is employed in conjunction with StyleGAN3 and Laplacian variance-based filtering to generate high-quality synthetic images, thereby increasing dataset diversity. To mitigate image distortion and edge aliasing during upsampling, a dynamic upsampling operator, DySample, is introduced. After deployment on the Jetson Xavier NX platform with TensorRT acceleration, the model achieves a mean precision of 89.24%, a mean recall of 87.71%, and a mAP@50 of 87.02%, running at 43 FPS. The overall performance surpasses the original YOLOv8n and other mainstream detection methods. This study provides a novel and practical solution for intelligent maturity monitoring of cotton and potentially other crops. The source code is publicly available at: https://github.com/mohenghui/Cotton_maturity .https://doi.org/10.1007/s44443-025-00081-8Maturity detectionDual backboneAttention mechanismUpsamplingTensorRT |
| spellingShingle | Meiqi Zhong Linjing Wei Henghui Mo YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection Journal of King Saud University: Computer and Information Sciences Maturity detection Dual backbone Attention mechanism Upsampling TensorRT |
| title | YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection |
| title_full | YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection |
| title_fullStr | YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection |
| title_full_unstemmed | YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection |
| title_short | YOLOv8n-DSDL: A lightweight dual-backbone network with decoupled self-attention for cotton maturity detection |
| title_sort | yolov8n dsdl a lightweight dual backbone network with decoupled self attention for cotton maturity detection |
| topic | Maturity detection Dual backbone Attention mechanism Upsampling TensorRT |
| url | https://doi.org/10.1007/s44443-025-00081-8 |
| work_keys_str_mv | AT meiqizhong yolov8ndsdlalightweightdualbackbonenetworkwithdecoupledselfattentionforcottonmaturitydetection AT linjingwei yolov8ndsdlalightweightdualbackbonenetworkwithdecoupledselfattentionforcottonmaturitydetection AT henghuimo yolov8ndsdlalightweightdualbackbonenetworkwithdecoupledselfattentionforcottonmaturitydetection |