An End-to-End Learning Architecture for Efficient Image Encoding and Deep Learning

Learning-based image/video codecs typically utilize the well known auto-encoder structure where the encoder transforms input data to a low-dimensional latent representation. Efficient latent encoding can reduce bandwidth needs during compression for transmission and storage. In this paper, we examine the effect of assigning high level coarse grouping labels to each latent vector. Designing coding profiles for each latent group can achieve high compression encoding. We show that such grouping can be learned via end-to-end optimization of the codec and the deep learning (DL) model to optimize rate-accuracy for a given data set. For cloud-based inference, source encoder can select a coding profile based on its learned grouping and encode the data features accordingly. Our test results on image classification show that significant performance improvement can be achieved with learned grouping over its non-grouping counterpart.