Denoising Hyperspectra with Diffusion Models

Hyperspectral imaging captures detailed information across narrow spectral bands for each pixel in an image, allowing us to identify objects and materials based on their unique spectral signatures. It has many important applications across a variety of industries, including the detection of cancerous tissues, monitoring soil to improve agriculture, and detecting contaminants in food. However, these images are often contaminated with noise due to hyperspectral sensor limitations and environmental conditions. Furthermore, these images are high-dimensional and very expensive to obtain, making hyperspectral image denoising a very complex process. To address these challenges, my approach focuses on denoising individual spectra rather than full hyperspectral images. By targeting spectral data at the pixel level, I'm developing a computationally efficient and accurate method to denoise, deblur, and super-resolve hyperspectral images by using Diffusion Models. I built a Denoising Diffusion Probabilistic Model (DDPM) to transform pure noise into spectra from the training distribution, by going through a reverse diffusion process that predicts and eliminates noise through time steps. I’m now leveraging the Denoising Diffusion Restoration Model (DDRM) technique to incorporate known, clean data to guide the reverse diffusion process, allowing me to restore a particular spectral wave rather than generating a random one from the original training distribution. My results have been very promising so far, especially at very high noise and at low resolution levels, where my DDRM implementation achieves substantially higher reconstruction scores than existing denoising methods.