Ising Machine-Based Similar Region Search Using Feature Histograms

We propose a novel algorithm for finding a similar region in images using Ising machines, which are a new type of non-von Neumann computers that specialize in solving combinatorial optimization problems. Applications of Ising machine are expected to be promising in diverse fields such as media processing and delivery. To leverage Ising machine-based approaches, a problem must be formulated in a quadratic unconstrained binary optimization (QUBO) form. In this study, to unleash the potentials of Ising machines in image processing, we present a method for formulating a similar region search as a QUBO problem. Experiments show that the QUBO-based algorithm realizes the detection of similar regions in accordance with policies for the similarity between images and the clumpiness of detected regions.

I. Introduction

Ising machines, a new type of non-von Neumann computing scheme specialized in solving combinatorial optimization problems, have attracted increasing attention recently. Since the first commercial quantum annealer emerged in 2011 [1], various types of Ising machines have been developed based on electronic and photonic systems as well as quantum systems [2]. A common feature of Ising machines is the minimization of the energy function of an Ising model or its equivalent; i.e., a quadratic unconstrained binary optimization (QUBO) form given by

$$\begin{equation*}H=\displaystyle \sum_{i}\sum_{j}q_{i,j^{X}i^{X}j},\ q_{i,j}\in \mathbb{R},\ x_{i}\in\{0,1\}. \tag{1}\end{equation*}$$