Rapid super resolution for infrared imagery.

Area of Science:

• Agricultural technology
• Computer vision
• Image processing

Background:

• Infrared (IR) imagery is crucial for agricultural applications like irrigation monitoring and disease detection.
• Existing IR cameras often suffer from low resolution, limiting their effectiveness in enhancing plant traits.
• Deep learning (DL) offers potential for image super-resolution but requires efficient computational methods.

Purpose of the Study:

• To develop a deep learning-based super-resolution method for low-power devices to enhance agricultural IR images.
• To improve the spatial resolution of IR imagery for better plant trait analysis.
• To create an efficient DL model with reduced computational complexity.

Main Methods:

• A novel deep learning approach was employed, performing most calculations in the low-resolution domain.
• Layer results were aggregated to facilitate information flow within the network.
• Depthwise separable convolution was utilized, achieving high performance with approximately 300K multiply-accumulate computations (MACs).

Main Results:

• The proposed method achieved significant super-resolution of IR images, enhancing plant traits.
• Achieved a ×4 improvement in spatial resolution, surpassing the diffraction limit, as confirmed by Modulation Transfer Function (MTF) analysis.
• Demonstrated superior efficiency compared to state-of-the-art methods, requiring substantially fewer MACs (300K vs. 1500K).

Conclusions:

• The developed DL method provides an effective and computationally efficient solution for IR image super-resolution in agriculture.
• The technique enhances the analysis of plant traits by improving image spatial resolution from low-power devices.
• This approach offers a practical advancement for precision agriculture through improved imaging capabilities.