Neural Data Compression with Content-Adaptive Latent Refinement

The patent performs preliminary optimization of the neural network parameters over a corpus of images during a training phase. This preliminary action establishes a baseline encoder and decoder that work well across multiple images, but then allows for content-adaptive refinement during actual compression by optimizing latent representations for each specific image, thus achieving both generalization and individual adaptation.

The patent introduces dynamics by allowing the latent representations to be adaptively optimized for each specific image content while keeping the neural network architecture fixed. This dynamic adjustment of latent variables based on individual image characteristics enables the system to achieve optimal rate-distortion performance for each image without requiring retraining of the entire network.

The patent extracts the content-adaptive optimization from the neural network training process itself. Instead of retraining the entire network for each image, it separates the global network parameter optimization (done once over a corpus) from the local latent representation optimization (done per image). This extraction allows efficient per-image adaptation without the computational burden of full network retraining.

The patent applies partial optimization by focusing computational effort only on optimizing the latent representations rather than the entire network. This partial action achieves the necessary adaptation for each image without the excessive computational cost of complete network retraining, balancing performance improvement with computational feasibility.