Experiments With Four Pattern Recognition Algorithms

Sammanfattning: A big part of computer vision concerns the issue ofhow well images can be classified into their corresponding classes.Image classification is a big part of reducing the gap betweenhuman and AI performance. Images are classified by first usinga dataset of images together with their given classifications totrain the system with machine learning. This can then be usedon images without any classification to test and see how well thealgorithm can classify an unknown image. However, this can bedone in many different ways and methods. The aim of the projectis to compare the performance of four algorithms: MultivariateGaussian Distribution Model, Least Square Regression (LS),Kernel Regression (KR) and Extreme Learning Machine (ELM).By using three different databases, the performance of the fouralgorithms varied. The algorithms were tested by reducing thedimensions of the images it was processing in order to analyzehow well the algorithms worked with different quality of images.The results show that some algorithms work best when moreimages are used for training, while some algorithms strugglewith more images and/or classes. This report concluded that theaccuracy of the algorithms varied depending on the structureof the database. By analyzing the structure of the database,the optimal classification algorithm can be chosen for bestperformance.