Features Denoising for Learned Image Coding

In recent years the advancements in the neural networks field have fostered the advent of end-to-end learned coding schemes capable of efficient image representations that reduce the required storage space and transmission time. In general, the features produced by these encoders are entropy-efficient and permit reconstructing the coded image with low distortion. However, whenever they are applied to a generic image, its latent representation might not be the optimal one in the feature space since the coding network parameters were trained to generalize on a wide set of images. This implies that reducing the distance between the optimal features and those generated by the network allows for better rate-distortion tradeoffs. This work proposes a feature denoising approach that permits improving the coding performance over five different learned image coding paradigms with minimal hyperparameters tuning. The proposed solution proved to be quite general and versatile since the decoding architecture and time are not affected by such optimization.