Multiscale spatial modeling with applications in image analysis

Abstract

Computer vision is a very important research area and is continuously growing. One of the prevalent research areas in computer vision is image matching. In image matching there are two main components, namely feature detection and feature matching. The aim of this this study is to determine whether Direct Sampling can be used for feature matching, and also if the combination of Direct Sampling and the Discrete Pulse Transform feature detector can be a successful image matching tool. In feature detection there are many strong methods including convolutional neural networks and scale-space models such as SIFT and SURF, which are very well-known feature detection algorithms. In this work we utilize another scale-space decomposition tool called the Discrete Pulse Transform (DPT). We particularly use the DPT decomposition to enable significant feature detection. We then concentrate on using the Direct Sampling algorithm, a stochastic spatial simulation algorithm, for modelling and matching of features. We do not consider convolutional neural networks or SIFT or SURF for texture matching in this work, this is because we particularly focus on the use of spatial statistics in image matching. We finally propose a novel multiscale spatial statistics feature detection and matching algorithm which combines the DPT feature detection with Direct Sampling for feature matching, specifically for texture classes of images. The performance of the proposed method is tested by comparing the distances obtained from the proposed algorithm between different texture images. We see that this proposed novel multiscale spatial modelling approach to feature matching with the focus on textures performs well at discriminating between difficult to discriminate between textures.