Lightweight rice leaf spot segmentation model based on improved DeepLabv3

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Rice is a vital food crop, but its susceptibility to diseases necessitates effective monitoring.
• Existing rice leaf spot segmentation models suffer from high computational overhead, hindering practical field application.
• Accurate segmentation is crucial for diagnosing diseases like rice blast, brown spot, and bacterial leaf blight.

Purpose of the Study:

• To develop a lightweight and efficient rice leaf spot segmentation model for field deployment.
• To improve segmentation accuracy, particularly in transitional regions and at lesion boundaries.
• To reduce computational complexity and model parameters for enhanced usability in resource-constrained environments.

Main Methods:

• Developed MMPC-DeepLabv3+, a lightweight model using MobileNetV3_Large (MV3L) as the backbone.
• Incorporated a multi-scale detail enhancement (MSDE) module with Haar wavelet downsampling for improved boundary and gap segmentation.
• Utilized a PagFm-Ghostconv Feature Fusion (PGFF) module with coordinate attention (CA) to reduce computational overhead and enhance robustness.
• Employed a hybrid loss function (Focal Loss + Dice Loss) to address class imbalance in disease imagery.

Main Results:

• MMPC-DeepLabv3+ achieved 81.23% mean intersection over union (MIoU) and 89.79% mean pixel accuracy (MPA) on natural illumination images.
• Significantly reduced computational resources: 9.695 G Flops and 3.556 M parameters.
• Outperformed the baseline DeepLabv3+ by 1.89% in MIoU and 0.83% in MPA, with 93.1% and 91.6% reductions in Flops and Params, respectively.

Conclusions:

• MMPC-DeepLabv3+ offers a superior balance of accuracy and computational efficiency compared to existing models like DeepLabv3+, U-Net, PSPNet, HRNetV2, and SegFormer.
• The model's lightweight design and high performance establish a new standard for rice lesion segmentation in precision agriculture.
• This research facilitates the practical application of AI for disease management in rice cultivation.