Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Lightweight Hyperspectral Image Super-Resolution with Depthwise Separable Dilated Convolutional Network

Usman Muhammad, Jorma Laaksonen, Lyudmila Mihaylova

TL;DR

This work tackles hyperspectral image super-resolution by introducing a lightweight depthwise separable dilated convolutional network (DSDCN) inspired by MobileNet, augmented with a dilated fusion block to capture multi-scale spatial-spectral features. A band-grouping strategy reduces spectral redundancy, while a tri-component loss combining $L_{MSE}$, $L_{SAM}$, and $L_{ ext{L2}}$ guides high-fidelity reconstruction. Empirical results on the PaviaC and PaviaU datasets show that DSDCN achieves competitive PSNR, MSSIM, and SAM scores with only ~0.96M parameters, outperforming several larger models on PaviaC and remaining highly competitive on PaviaU. Overall, the method offers a practical, resource-efficient solution for hyperspectral SR suitable for real-world applications.

Abstract

Deep neural networks have demonstrated highly competitive performance in super-resolution (SR) for natural images by learning mappings from low-resolution (LR) to high-resolution (HR) images. However, hyperspectral super-resolution remains an ill-posed problem due to the high spectral dimensionality of the data and the scarcity of available training samples. Moreover, existing methods often rely on large models with a high number of parameters or require the fusion with panchromatic or RGB images, both of which are often impractical in real-world scenarios. Inspired by the MobileNet architecture, we introduce a lightweight depthwise separable dilated convolutional network (DSDCN) to address the aforementioned challenges. Specifically, our model leverages multiple depthwise separable convolutions, similar to the MobileNet architecture, and further incorporates a dilated convolution fusion block to make the model more flexible for the extraction of both spatial and spectral features. In addition, we propose a custom loss function that combines mean squared error (MSE), an L2 norm regularization-based constraint, and a spectral angle-based loss, ensuring the preservation of both spectral and spatial details. The proposed model achieves very competitive performance on two publicly available hyperspectral datasets, making it well-suited for hyperspectral image super-resolution tasks. The source codes are publicly available at: \href{https://github.com/Usman1021/lightweight}{https://github.com/Usman1021/lightweight}.

Towards Lightweight Hyperspectral Image Super-Resolution with Depthwise Separable Dilated Convolutional Network

TL;DR

This work tackles hyperspectral image super-resolution by introducing a lightweight depthwise separable dilated convolutional network (DSDCN) inspired by MobileNet, augmented with a dilated fusion block to capture multi-scale spatial-spectral features. A band-grouping strategy reduces spectral redundancy, while a tri-component loss combining , , and guides high-fidelity reconstruction. Empirical results on the PaviaC and PaviaU datasets show that DSDCN achieves competitive PSNR, MSSIM, and SAM scores with only ~0.96M parameters, outperforming several larger models on PaviaC and remaining highly competitive on PaviaU. Overall, the method offers a practical, resource-efficient solution for hyperspectral SR suitable for real-world applications.

Abstract

Deep neural networks have demonstrated highly competitive performance in super-resolution (SR) for natural images by learning mappings from low-resolution (LR) to high-resolution (HR) images. However, hyperspectral super-resolution remains an ill-posed problem due to the high spectral dimensionality of the data and the scarcity of available training samples. Moreover, existing methods often rely on large models with a high number of parameters or require the fusion with panchromatic or RGB images, both of which are often impractical in real-world scenarios. Inspired by the MobileNet architecture, we introduce a lightweight depthwise separable dilated convolutional network (DSDCN) to address the aforementioned challenges. Specifically, our model leverages multiple depthwise separable convolutions, similar to the MobileNet architecture, and further incorporates a dilated convolution fusion block to make the model more flexible for the extraction of both spatial and spectral features. In addition, we propose a custom loss function that combines mean squared error (MSE), an L2 norm regularization-based constraint, and a spectral angle-based loss, ensuring the preservation of both spectral and spatial details. The proposed model achieves very competitive performance on two publicly available hyperspectral datasets, making it well-suited for hyperspectral image super-resolution tasks. The source codes are publicly available at: \href{https://github.com/Usman1021/lightweight}{https://github.com/Usman1021/lightweight}.
Paper Structure (12 sections, 12 equations, 1 figure, 2 tables)

This paper contains 12 sections, 12 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Band Grouping
  4. Depthwise Separable Convolution
  5. Dilated Convolution Fusion
  6. Upsampling with Transpose Convolution
  7. Custom Loss Function
  8. Experimental setup
  9. Datasets and Implementation
  10. Ablation Study on Band Grouping
  11. Comparison with State-of-the-Art Methods
  12. Conclusion

Figures (1)

  • Figure 1: A detailed overview of the proposed model. Shallow features are initially extracted using a lightweight separable convolutional layer with ReLU activation. The gray color represent depthwise convolutional blocks for deep feature extraction. The blue block denotes the dilated fusion module, capturing spatial and spectral features in parallel. The pink block performs upsampling to reconstruct the high-resolution output.