Hello all,
Would anyone here who has a PhD degree be interested in being part of the jury for my PhD defense (most likely in January)? My thesis is about image denoising and compression using neural networks.
This is pretty unconventional but my main motivation has been to make really effective and widely applicable image denoising using neural networks for my use-case which is image development using exclusively open-source software (and ideally I would like my research to be integrated into such software), so I think I am likely to find the best target audience here.
Below is the abstract:
This thesis addresses two fundamental challenges in image processing—image compression and image denoising—using deep learning techniques to improve both visual quality and computational efficiency.
The first challenge is image compression, which aims to minimize the storage and transmission cost of images while preserving visual quality. Recent methods based on convolutional autoencoders have shown superior results, but their reliance on complex entropy models for predicting the probability distribution of each feature leads to higher computational costs. This work proposes a simplified compression scheme that uses a single convolutional autoencoder with a set of multiple learned prior distributions stored in static tables of cumulative distribution functions, as an alternative to computationally expensive single-feature parametric priors. During inference, these static priors allow the entropy coder to efficiently compress spatial features across all channels. The proposed method achieves comparable rate-distortion performance to other state-of-the-art models, while significantly reducing entropy coding and decoding complexity.
The second challenge, image denoising, involves the removal of unwanted noise from images captured under suboptimal conditions. Noise not only degrades image quality but also impairs the performance of both standard and learned compression algorithms, as it is inherently non-compressible. This thesis proposes a unified model that performs joint denoising and compression. By training the model on noisy-clean image pairs across a wide range of noise levels, it learns to denoise images effectively as part of the compression process, all while maintaining the computational cost of compression alone. This joint approach improves rate-distortion performance compared to compressing noisy images or using separate denoising and compression models. Additionally, the model is capable of producing decompressed images with visual quality superior to that of the noisy uncompressed inputs.
The final part of this thesis focuses on raw input images, specifically Bayer pattern images produced by most digital cameras. It is demonstrated that processing raw or minimally processed images (e.g., debayered and converted to the standard linear Rec. 2020 color profile) offers substantial gains in both compression efficiency and denoising quality compared to working with fully processed sRGB images. Treating Bayer images as 4-channel inputs reduces computational complexity by a factor of four, while also improving compression performance at lower bitrates. Moreover, denoising raw or linear RGB images early in the processing pipeline enables greater generalization. Unlike models trained on processed sRGB images, which perform well only on data processed by specific camera image signal processors (ISPs) or software pipelines, early-stage denoising of raw or linear RGB data ensures compatibility across diverse imaging systems, development software, and processing styles.
To facilitate further research and development, a novel dataset of raw clean-noisy image pairs is introduced, with each scene containing at least one ground-truth clean image paired with multiple noisy versions captured at varying ISO levels and exposure times. This dataset supports academic research and the development of denoising models integrated into the pixel pipeline of image processing software, providing a valuable resource for advancing denoising and compression techniques in real-world applications.
List of publications:
[0] B. Brummer and C. De Vleeschouwer, Natural image noise dataset, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019 (as part of my master thesis work)
[1] B. Brummer and C. De Vleeschouwer, Adapting jpeg xs gains and priorities to tasks and contents, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2020, pp. 629–633.
[2] B. Brummer and C. De Vleeschouwer, End-to-end optimized image compression with competition of prior distributions, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2021, pp. 1890–1894.
[3] B. Brummer and C. De Vleeschouwer, On the importance of denoising when learning to compress images, in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), January 2023, pp. 2440–2448.
[4] B. Brummer and C. De Vleeschouwer, Joint learned compression and denoising of raw images, ongoing work
