Lightweight detection of cotton leaf diseases using StyleGAN2-ADA and decoupled focused self-attention
Abstract In cotton cultivation, diseases such as leaf spot, helminthosporium leaf spot, fusarium wilt, boll gray mold, and leaf curl significantly affect yield and quality. Current models face challenges like diverse disease traits, variable stages, small target detection, uneven lighting, and occlu...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-05-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00054-x |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract In cotton cultivation, diseases such as leaf spot, helminthosporium leaf spot, fusarium wilt, boll gray mold, and leaf curl significantly affect yield and quality. Current models face challenges like diverse disease traits, variable stages, small target detection, uneven lighting, and occlusions, resulting in low accuracy and adaptability. This study introduces the RT-DETR-DFSA model for cotton leaf disease detection, building on the Real-Time Detection Transformer (RT-DETR). The Decoupled Focused Self-Attention (DFSA) mechanism splits traditional two-dimensional self-attention into one-dimensional operations that are processed by a dilated convolution layer, merges positional features with the original input, enhances feature relationships, and dynamically adjusts self-attention weights. Style Generative Adversarial Network with Adaptive Discriminator Augmentation (StyleGAN2-ADA) and Fourier Transform are used to generate realistic images of cotton diseases, enhancing training and validation sets. The Layer-Adaptive Magnitude-based Pruning (LAMP) method reduces computational and memory demands, and the teacher-assistant-student architecture further improves accuracy through knowledge distillation. The RT-DETR-DFSA model achieves 87.14% detection accuracy, 85.03% recall, and 86.33% mean Average Precision (mAP50). Post-pruning, the model’s parameters are reduced to 4.9 million (M), with a computational demand of 31.5 Giga Floating-Point Operations Per Second (GFLOPs), showing superior performance over existing models. This provides technical support for cotton monitoring and insights for disease detection in other crops. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |