Assessing the prospects for robust sub-diffraction limited super-resolution imaging with deep neural networks

摘要

High resolution images are critical for a wide variety of military detection, recognition and identification tasks. Super-resolution reconstruction algorithms aim to enhance the image resolution beyond the capability of the imaging system being used. Until recently, undersampling of the optical signal on the image sensor has been the key factor limiting the attainable resolution of visible and infrared imaging systems. Traditional SR algorithms aim to overcome this undersampling by combining data from multiple frames in a sequence. However, recent advances in manufacturing technologies have led to a steady increase in the number of pixels in an image sensor. Instead, image blur caused by optical diffraction is becoming an important limitation to the attainable image resolution. Here we investigate if image resolutions beyond the limitations posed by optical diffraction may be achieved using deep neural network based single image super-resolution algorithms. These networks learn a mapping from low resolution images to high resolution counterparts from pairs of training images. This could allow them to reconstruct high frequency information beyond the diffraction limit based on prior information about likely scene contents. We find that an average gain in image resolution of over 30% could be achieved by such networks on simulated diffraction limited imagery. In addition we investigate how robust these networks are to the presence of noise in the low resolution input imagery. We show that low noise levels can lead to poor reconstruction results with networks trained on noise free examples, but also that training on multiple noise levels can be used to mitigate this deterioration in performance.