The Application of Harmony Search in Computer Vision

Abstract

The harmony search algorithm was developed in 2001 as a heuristic optimisation algorithm for use in diverse optimisation problems. After its introduction it was extensively used in multiple engineering disciplines with great success. In order to demonstrate the value of harmony search in computer vision applications I developed four novel algorithms based on harmony search that efficiently solves three problems that are commonly found in computer vision, namely visual tracking, visual correspondence matching and binary image restoration. Computer vision is a large discipline that includes solving many different kinds of optimisation problems. Many of these optimisation problems are discontinuous with derivative information difficult or impossible to come by. The most common solution is to use population based statistical optimisation algorithms like the particle filter, genetic algorithms, PSO, etc. but harmony search has never been investigated as a possible alternative. This is surprising since harmony search has been shown to be superior to these methods in several other engineering disciplines.

I therefore aim to show that harmony search deserves to be included in the computer vision researcher's toolbox of optimisation algorithms through the introduction of four novel algorithms based on harmony search that solve three diverse problems in computer vision. First the harmony filter (HF) is introduced as a visual tracking algorithm that is shown to be superior to the particle filter and the unscented Kalman filter (UKF) in both speed and accuracy for robust tracking in challenging situations. The directed correspondence search (DCS) algorithm is then introduced as a solution to the visual correspondence problem. Finally, two algorithms, counterpoint harmony search (CHS) and largest error first harmony search (LEFHS), are introduced for the blind deconvolution of binary images.

Comparative results from these algorithms are very promising. The harmony filter was compared with the particle filter and the UKF both of which have been extensively used in visual tracking. In challenging situations consisting of rapid and erratic target movement, extended periods of total and partial occlusion and changing light conditions, the HF proved to be more accurate and faster on average than both the particle filter and the UKF. Under various conditions I show that the HF is at least 2 times faster than a UKF implementation and 4 times faster than a particle filter implementation (using 300 particles).

While there are fewer algorithms specialising in the blind deconvolution of binary images, CHS and LEFHS were compared with a current state-of-the-art method and proved to be more robust to noise and more accurate. LEFHS is the only algorithm currently available that can recover a 24 x 12 binary image using blind deconvolution to 100% accuracy without putting constraints on the point spread function (blurring kernel).

During the development of these algorithms several valuable insights into the inner workings of harmony search were discovered. In each application harmony search had to be adapted in a different way and with each new adaptation a deeper understanding of the advantages of harmony search is revealed. Knowing which components may be modified without degrading performance is key to adapting harmony search for use in diverse problems and allows one to use harmony search in situations it was not originally designed for without losing its superior performance. These insights and the adaptation strategies that they lead to are the main contribution of this thesis.