Hyperspectral image compression using implicit neural representations

Hyperspectral images (HSI) capture the full electromagnetic spectrum for each pixel in a scene. They often hold hundreds of channels per pixel, providing significantly more information compared to a comparably sized RGB color image. As the cost of obtaining these images decreases, there is a need to create effective ways for storing, transferring, and interpreting hyperspectral data. In this thesis, we develop several neural compression-based methods for hyperspectral images. Our methodology relies on transforming hyperspectral images into implicit neural representations (INR), specifically neural functions that establish a correspondence between coordinates and features. We use a multilayer perception (MLP) network with sinusoidal activation functions that “learns” to map pixel locations to pixel spectrum for a given hyperspectral image. This representation thus acts as a compressed encoding of this image, and the original image is reconstructed by evaluating this network at each pixel location. In the other variation of using implicit neural representation to compress hyperspectral images, a sampling scheme is introduced to achieve better compression times while keeping decoding errors low. In our other method, instead of explicitly saving the weights of the implicit neural representation, the modulations that are applied to a base network that has been meta-learned are recorded. These modulations serve as a compressed coding for the hyperspectral image. An assessment of the proposed approach was conducted using four benchmarks: Indian Pines, Jasper Ridge, Pavia University, and Cuprite. The proposed method is evaluated against sixteen other schemes ((1) JPEG, (2) JPEG2000, (3) PCA-DCT, (4) PCA-JPEG2000, (5) MPEG, (6) X264, (7) X265, (8) PCA-X264, (9) PCA-X265, (10) FPCA-JPEG2000, (11) 3D-DCT, (12) 3D-DWT-SVR, (13) WSRC, (14) HEVC, (15) RPM, and (16) 3D-SPECK.) for hyperspectral image compression, and according to the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics, the method developed in this study achieves state-of-the-art compression rates at low-bit rates. We also used a large hyperspectral image dataset, compressed it using our methods, and compared our results with JPEG and MPEG. Finally, we conducted task-aware hyperspectral image compression, in which regions are chosen according to a task, and hyperspectral images are compressed using our proposed method.

Name thesis:degree_name

Doctor of Philosophy (PhD)

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Ontario Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Rezasoltani, Shima

Advisor dc:contributor.advisor

• Qureshi, Faisal

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10155/1993

OAI identifier oai:identifier

oai:ontariotechu.scholaris.ca:10155/1993