Automated Video Quality Assessment for Deep-Sea Video

Abstract:

Video provides a rich source of data for geophysical analysis, often supplying detailed information about the environment when other instruments may not. This is especially true of deep-sea environments, where direct visual observations cannot be made. As computer vision techniques improve and volumes of video data increase, automated video analysis is emerging as a practical alternative to labor-intensive manual analysis. Automated techniques can be much more sensitive to video quality than their manual counterparts, so performing quality assessment before doing full analysis is critical to producing valid results.

Ocean Networks Canada (ONC), an initiative of the University of Victoria, operates cabled ocean observatories that supply continuous power and Internet connectivity to a broad suite of subsea instruments from the coast to the deep sea, including video and still cameras. This network of ocean observatories has produced almost 20,000 hours of video (about 38 hours are recorded each day) and an additional 8,000 hours of logs from remotely operated vehicle (ROV) dives.

We begin by surveying some ways in which deep-sea video poses challenges for automated analysis, including: 1. Non-uniform lighting: Single, directional, light sources produce uneven luminance distributions and shadows; remotely operated lighting equipment are also susceptible to technical failures. 2. Particulate noise: Turbidity and marine snow are often present in underwater video; particles in the water column can have sharper focus and higher contrast than the objects of interest due to their proximity to the light source and can also influence the camera's autofocus and auto white-balance routines. 3. Color distortion (low contrast): The rate of absorption of light in water varies by wavelength, and is higher overall than in air, altering apparent colors and lowering the contrast of objects at a distance.

We also describe measures under development at ONC for detecting and mitigating these effects. These steps include filtering out unusable data, color and luminance balancing, and choosing the most appropriate image descriptors. We apply these techniques to generate automated quality assessment of video data and illustrate their utility with an example application where we perform vision-based substrate classification.