Compressed Sensing for Computer Vision

点击翻译按钮获取中文摘要

The goal of computer vision (CV) is automated visual recognition by processing images and videos. CV significantly lags behind human visual capability. However, a few recent success stories in commercial applications are noteworthy: face detection and recognition, optical character recognition, Kinect gaming, fingerprint verification and so on. Thus, ample research opportunity exists in the progression of CV in its theory and practice. The principal roadblock to the success of computer vision research seem to be the lack of a suitable numerical representation of images that the computer algorithms (AI / machine learning) can easily process. In particular, an image could be a collection of a million or more pixels, which is overwhelming for an algorithm to process meaningfully. The art will be representing these pixels in a compressed form, which would be useful for the algorithms. In this research program, I propose to look at compressed sensing (CS)-based representation of images for CV. CS is a rapidly emerging area in signal processing research that has acted as a paradigm shift in signal processing applications, because using CS, some signals can be reconstructed using far fewer samples than what the classical theory predicts. A prominent application example of CS is rapid magnetic resonance imaging, where the time of signal acquisition has been significantly reduced. In this research proposal, I explain the rational behind using CS within CV applications. Potential impact of the proposed research is the proliferation of CV methods to limited memory devices, such as wearable devices and understanding of the image representation for CV. The program also plans to train 4 PhD and 2 MSc students in the state-of-the-art CV research.