Home // SIGNAL 2022, The Seventh International Conference on Advances in Signal, Image and Video Processing // View article

STP-Net: Semi-Tensor Product Neural Network for Image Compressive Sensing

Authors:
Youhao Yu
Richard M Dansereau

Keywords: compressive sensing; convolutional neural network; semi-tensor product; image reconstruction.

Abstract:
Semi-tensor product (STP) is developed into a neural network in this paper and applied to image compressive sensing (CS). Large matrix computation for fully connected layers results in a large number of weight coefficients that need long training times. Instead of using an M×N measurement matrix, according to the theory of STP a smaller measurement matrix of size M/t×N/t can be applied, where t is a shrinkage factor. STP only needs N/t elements of the original signal for one measurement and the measurement matrix is shrunk to 1/t^2 that of traditional CS. The shrinkage factor t is adjustable. To demonstrate the effectiveness of the STP-based neural network, we apply it to image reconstruction. The goal is to sample and recover larger images, without partitioning into smaller blocks that introduces block artifacts, and provide good initial reconstruction for subsequent networks.

Pages: 7 to 12

Copyright: Copyright (c) IARIA, 2022

Publication date: May 22, 2022

Published in: conference

ISSN: 2519-8432

ISBN: 978-1-61208-970-6

Location: Venice, Italy

Dates: from May 22, 2022 to May 26, 2022