Paddy pest image segmentation based on multiscale attention fusion VM-UNet

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Accurate paddy pest image segmentation (PPIS) is crucial for smart agriculture but challenging in natural environments.
• Existing Convolutional Neural Networks (CNNs) and Transformers struggle with global dependencies and computational complexity, respectively.

Purpose of the Study:

• To develop a novel multiscale attention fusion VM-UNet (MSAF-VMUNet) for precise PPIS.
• To improve the modeling of global and local dependencies for enhanced pest detection.

Main Methods:

• The MSAF-VMUNet integrates Visual State Space Models (VSS) for long-range dependencies and U-Net for precise localization.
• Multiscale VSS (MSVSS) blocks capture contextual information, while an improved attention fusion (IAF) module aids multi-level feature learning.
• An Attention VSS module in the bottleneck adaptively refines feature emphasis.

Main Results:

• MSAF-VMUNet effectively models global-local relationships and context across scales without increasing computational load.
• The model achieved a PPIS precision of 79.17% on the IP102 dataset, outperforming U-Net by 15.51% and VM-UNet by 3.39%.
• It successfully addressed challenges like small pest detection, occlusion, noise, and preprocessing needs.

Conclusions:

• MSAF-VMUNet offers an effective and reliable solution for real-time PPIS in smart agriculture.
• The proposed architecture enhances pest control detection systems by improving segmentation accuracy and robustness.
• This research contributes to advancing automated pest management through deep learning.