SBIR Phase I:Improved image compression targeting machine learning based detection algorithms

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is in reducing the sizes of earth observation images, thereby saving millions of dollars on transmission, storage and processing costs. Storage and processing costs account for a large portion of costs for earth observation companies, reducing these costs will make earth observation imagery more accessible for a broader audience including companies performing climate and environment studies. The need for reduced image sizes also extends to video compression. Video calls, telepresence, telemedicine, remote work and metaverse all depend on the ability to stream video over limited or variable bandwidth connections. This technology will find ready applications in the area of high-efficiency video compression._x000D_ _x000D_ This SBIR Phase I project uses object detection algorithms to detect areas of importance in an image and utilizes that information to improve the efficiency of image compression. More specifically, the research will develop a compression network that maximizes the detection accuracy of a down-stream, machine learning-based object detector. In contrast, current compression algorithms do not interpret the images they are compressing and simply minimize a visual loss function that treats the entire image equally. The technology will produce images that can be stored in the standard image compression file formats including .png and .jpeg. This technology will enable fast compression and will explore modified compression architectures, quantization, pruning and parallelization using graphics processing units to reduce latency of compression._x000D_ _x000D_ This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.