End-to-End Image Compression Through Machine Semantics

With the increasing demand for AI automated analysis, machine semantics have replaced signals as a new focus in visual information compression. In this paper, we propose a novel end-to-end machine semantic information compression method. To better align machine semantics with tasks, we jointly optimize the entire process of semantic extraction, compression, and inference. Additionally, we introduce a dedicated activation function called StairReLU for machine semantic compression. StairReLU considers the non-linearity in neural networks and quantization in data compression as the same problem, enabling machines to adapt to semantic representations with lower information entropy during the training process. Finally, we conduct end-to-end machine semantic compression experiments on three different datasets. The results demonstrate the superiority of our proposed approach over traditional compression coding methods and incomplete end-to-end compression methods in terms of the trade-off between bit rate and quality.