The Image Semantic Compression Method for Underwater Robotic Applications

Abstract

With significant progress in developing intervention class AUVs, the visual control of manipulation tasks has become particularly relevant. Due to low bandwidth acoustic communication, it is necessary to seek new image compression and data transportation methods. This paper presents a novel method (Image Semantic Compression–ISC) for image compression based on semantic segmentation of camera images followed by semantic objects simplification and binary compression. The description of the proposed method and its comparative study with different image compression methods and different binary compression algorithms (Deflate, bzip2, LZMA, Brotli) are presented. The comparative study is conducted on the dataset of ~1500 different underwater images where the SUIM-Net model was trained. Results of framerate evaluation for different distances and frame resolution are presented in the article. It is shown that a threefold framerate increase with the frame resolution 320$\times $256(1.48 FPS vs. 0.51 FPS on a distance of 200 meters) can be achieved due to semantic image compression in comparison with low-quality image compression based on the JPEG algorithm.