Current Frontiers In Neural Image Compression: The Rate-Distortion-Computation Trade-Off And Optimizing For Subjective Visual Quality

Learning-based methods for image and video compression hold the promise of significant advances in rate-distortion performance. To date, however, neural image compression has largely underdelivered relative to traditional measures of image quality despite using considerably more computation for decoding. Optimizing explicitly for alternative quality metrics that better correlate with human preferences leads to significant rate savings over standard codecs, but the resulting reconstructions often see little benefit on subjective evaluation tests. Similarly, smaller models can improve decode speed, but they typically lead to a commensurate drop-in rate-distortion performance. This talk will cover state of the art neural compression models along with two crucial research directions for learning-based compression. First, weƒ??ll explore the rate-distortion-computation trade-off across different architectures to better understand how neural methods might be able to achieve the decode speed required by typical applications. And, second, weƒ??ll discuss recent research that combines perceptual metrics and adversarial methods to boost subjective quality and provide a nearly 50% rate savings over standard codecs. Conducting these surveys has required a tremendous multidisciplinary effort, from building the software and hardware sensing systems, to collecting the data, and finally to producing data products for statisticians, biologists, and the further development of machine learning models. This talk will give an overview of how we utilize machine learning in our surveys, discuss some of the unique challenges that come with developing and deploying these technologies in remote sensing research applications, and share how weƒ??ve addressed these challenges.