G-RRDB: An Effective THz Image-Denoising Model for Moldy Wheat

Area of Science:

• Agricultural Science
• Image Processing
• Computer Vision

Background:

• Terahertz (THz) imaging of wheat is susceptible to noise and feature degradation due to light source fluctuations.
• Effective denoising is crucial for accurate feature extraction and subsequent analysis, such as disease identification.

Purpose of the Study:

• To propose a novel THz image denoising model, G-RRDB, for wheat.
• To enhance the global sensory field and feature attention capabilities of the denoising network.
• To improve the accuracy of moldy wheat classification using denoised THz images.

Main Methods:

• Development of the Ghost-LKA module, combining large kernel convolutional attention with Ghost convolution for a global sensory field.
• Integration of spatial and channel attention into the DAB module to enhance feature focus.
• Construction of the G-RRDB model by integrating Ghost-LKA and DAB modules with a dense residual baseline.
• Evaluation using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Main Results:

• The G-RRDB model demonstrated improved PSNR and SSIM compared to traditional denoising networks.
• Classification accuracy of moldy wheat using VGG16 reached 92.8% with G-RRDB denoised images.
• This represents a 1.7% and 0.2% improvement over baseline and DAB-enhanced baseline models, respectively.

Conclusions:

• The proposed G-RRDB model exhibits excellent denoising performance for THz images of wheat.
• The integration of Ghost-LKA and DAB modules effectively addresses noise and enhances feature representation.
• G-RRDB facilitates more accurate classification of moldy wheat, showcasing its practical utility.