System and method for designing super resolution deep convolutional neural networks

Abstract

Apparatuses and methods of manufacturing same, systems, and methods for generating a convolutional neural network (CNN) are described. In one aspect, a minimal CNN having, e.g., three or more layers is trained. Cascade training may be performed on the trained CNN to insert one or more intermediate layers until a training error is less than a threshold. When cascade training is complete, cascade network trimming of the CNN output from the cascade training may be performed to improve computational efficiency. To further reduce network parameters, convolutional filters may be replaced with dilated convolutional filters with the same receptive field, followed by additional training / fine-tuning.

Description

[0001] Cross References to Related Applications [0002] This application claims U.S. Provisional Patent Application Serial No. 62 / 471,816, filed March 15, 2017, and U.S. Nonprovisional Patent Application Serial No. 15 / 655,557, filed July 20, 2017 Priority, said application is hereby incorporated by reference in its entirety. technical field [0003] The present disclosure relates generally to image super-resolution, and more specifically, to systems and methods for designing efficient super-resolution deep convolutional neural networks through cascaded network training, cascaded network pruning, and dilated convolutions. Background technique [0004] Super-resolution imaging generates high-resolution (HR) images from low-resolution (LR) images. Super-resolution (SR) imaging has wide applicability, from surveillance and face / iris recognition to medical image processing, and directly increases the resolution of images and videos. Many algorithms / systems for performing SR ha...

AI Technical Summary

Problems solved by technology

In response, researchers have proposed increasing the size of the SRCNN, but many proposals use extremely large numbers of parameters, and many of the SRCNNs under discussion cannot perform in real time.

Guessing the appropriate training settings (i.e., learning rate, weight initialization, and weight decay) can be very difficult due to the proposed large network size

As a result, training may not converge at all or fall into a local minimum

Images