|Project Title:||Exploiting Video Analytics to Select Polarization Imaging Information for Ensuring Video Content Clarity|
|Hosting Institution:||LSCM R&D Centre (LSCM)|
|Abstract:||What can be seen or captured by camera highly depends on light conditions at that moment. For example, visual content beneath shinny surface may be partly blocked by other light sources reflected from there. The invention of polarized lenses in 1936 has helped photographers take crystal clear pictures by cutting hazardous glare off of flat surfaces such as water, glass, and asphalt. Nonetheless, the commonly used circular polarizer adding to a lens must be manually turned to block off unwanted light oscillating in angles other than the chosen angle. The light not being blocked off is polarized light for forming the clearest visual content. The circular polarizer is ideal for taking still pictures. However, human evaluation of whether content clarity is achieved is a critical factor.
This project attempts to apply video analytics techniques on raw visual content embedded with polarization imaging information to enable automatic capture of content-specific visual imagery in continuous manner. One of the content-specific visual imagery is to clearly see fish or activities beneath water surface in a pond by removing unwanted imagery of sky or nearby buildings reflecting from rippling waves. Another content-specific visual imagery is to unveil internal stress of transparent plastics and glass that is invisible with conventional imaging techniques. The project will explore automated means for selecting appropriate polarized visual data from convoluted light sources to clearly reveal content-specific items. The work will leverage on the new polarization image sensor technology for continuously capture video streams free from unwanted visual content due to ever-changing light sources. The R&D work focuses on the investigation of how video analytics techniques can be applied to discover or assist to select content of interest and subsequently detect similar content items from polarized video content. Deep learning techniques will be involved in the R&D methodology. The R&D results may help the deployment of various video analytics applications under ever-changing light conditions by providing clear video content for detecting and tracking objects of interest without any human intervention.
|Project Coordinator:||Dr Dorbin Ng|
|Approved Funding Amount:||HK$2.79M|
|Project Period:||31 Mar 2020 - 30 Mar 2021|