Semantic image compression based on data hiding

This study proposes a novel scheme of semantic image compression. A compressor firstly creates a compact image by gathering a part of pixels in an original image, and calculates estimation errors of the rest pixels. Then, a compressed image is produced by embedding the estimation errors into the compact image using data hiding techniques. This way, the compressed image are made up of a small number of pixel values, and the original content is still visible roughly through the compressed image without any decompression tool. If a decompression tool is available, a user may reconstruct a high quality image with original size by exploiting the embedded data. Because the proposed scheme is compatible with reversible and non-reversible data hiding techniques, either the lossy or lossless semantic compression can be performed. With different parameters, the qualities of compressed and decompressed images vary. Furthermore, the smoother the original image content, the better is the compression-decompression performance.