Underwater Forward-Scan Sonar Video Coding by Background Modeling and Synthesis for Real-Time Transmission

The traditional navigation cameras are being replaced by 2-D forward-scan sonar video devices, when untethered submersible robots operate in underwater environments with poor visibility. A typical FS sonar image comprises of object highlight and cast shadow regions, as well as large relatively uniform background areas, generally corrupted with high level of speckle noise. While both are important in image analysis for object and scene interpretation, highlight and shadow regions encode valuable information for inferring 3-D shape. For some tasks, real-time man-machine interaction and cooperation can significantly enhance the robot performance. The transmission of live video from the robotic platform to a remote surface station requires high video compression ratios to meet the current low-bandwidth limitations of acoustic channels. This work makes use of a novel forward-scan sonar image coding scheme, where 1) object highlights and shadow edges (treated as foreground) are segmented from the less informative background, maintaining high details for encoding; 2) background is highly compressed to transmit at very little cost; 3) together, they are decoded at the receiver end for reconstruction. Application to various video sequences shows an average compression ratio of 100, for raw video data at 512×96 resolution transmitted at 10 frame per second.