Joint learned compression and denoising schemes for improved photographic image quality and reduced processing complexity

This thesis addresses image compression and denoising using deep learning, enhancing visual quality and computational efficiency. For compression, convolutional autoencoders are improved by replacing computationally expensive, single-feature parametric entropy models with a simplified scheme using multiple, pre-learned, static cumulative distribution function tables. This significantly reduces entropy coding/decoding complexity during inference. 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 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 as part of the compression process 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 input. The final part of this thesis focuses on raw input images. Processing raw or minimally processed images offers substantial gains in both compression efficiency and denoising quality compared to working with fully processed images. Treating Bayer images as 4-channel inputs reduces the computational complexity of denoising models and compression encoders 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. A novel dataset of raw cleannoisy image pairs is introduced to support further research and the development of models integrated into image processing software pipelines.