Super-resolution reconstruction and denoising of 3D millimetre-wave images using a complex-valued convolutional neural network

Imaging systems leveraging millimetre-wave (mmW) frequencies have several advantages, however, such systems suffer from poor resolution images as compared to higher frequency reconstructions such as in optical regime. Also, practical radar systems are susceptible to noise such as clutter, thermal noise, motion blurs, etc. To recover the original mmW image from these poorly resolved noisy images, two individual image processing steps are required, that is, super-resolution and denoising. This paper focuses on using a complex-valued convolutional neural network (CV-CNN) to combine the two individual processing steps into one single algorithm. By designing the CV-CNN to accommodate complex-valued reconstruction data, the phase information content of the input images, along with the magnitude information, is considered in the process. A computational imaging (CI) numerical model, instead of an experimental imaging system, is used to train and test the neural network. By comparing the performance metrics of the final reconstruction images, it is observed that the developed CV-CNN can resolve and de-noise the poorly resolved noisy input mmW images to a high degree of fidelity.