Subject

Hyperspectral images (HS) such as PRISMA and EnMAP are state-of-the-art spaceborne missions for Earth Observation with rich spectral information acquired at 400-2500nm with 200 spectral bands. One weakness of such powerful images is their coarse spatial resolution. Over last decades, focused research and studies have been made to improve their spatial resolution. A popular and promising approach is through image fusion. Using a HD multi-spectral image such as Worldview 3 with over a dozen spectral bands at 2-4m, methods have been developed to enhance the low-resolution HS images. The result is a HS image cube at 2m resolution [1]. To operationalize the algorithms in real case scenarios, the ultimate goal is to implement the fusion on hardware.

Kind of work

The objective of this thesis is to accelerate an existing CNN-based hyperspectral image fusion algorithms on a hardware accelerator. It can be either on an FPGAs using high-level synthesis tools (using high-programming languages (C/C++) for the architecture description [2]) or using Tensor Processing Units such as Google CORAL.

Number of Students

2

Expected Student Profile

• Knowledge of Matlab and C/C++.
• Interested in image processing, hardware acceleration or FPGAs.


References:
[1] Yang, J., Zhao, Y. Q., & Chan, J. C. W. (2018). Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5), 800.
[2] UG902, X. (2021). Vivado Design Suite User Guide, High-Level Synthesis.