University of Joensuu Pattern Recognition
Overview
Here are applets which demonstrate different algorithms.


BAYESIAN CLASSIFIER: Applet will focus on two-class case in which the user can define both a priori probabilities and the class-conditional probability density functions for each of the classes. The user can choose whether to minimize the classification error probability or the average risk.

kNN CLASSIFIER: Applet will focus on two-class case in which k nearest neighbors are identified for a given feature vector (data point) with respect to the euclidean distance. The feature vector is classified to the class with the maximum number of samples in k.

LINEAR CLASSIFIER: Applet is based on the assumption that all feature vectors (data points) from the two available classes can be classified correctly by using a linear discriminant function. The algorithm is initialized by an arbitrary line, which is corrected systematically until all points are corectly classified. converges.

BSAS: In Basic Sequenial Algorithmic Scheme applet the first feature vector forms the first cluster and after that each new feature vector is either assigned to an existing cluster or assigned to form a new cluster. The decision is based on the threshold value and the maximum number of clusters, which are defined by the user beforehand.

GAS: In Generalized Agglomerative Scheme the initialization is performed by assigning each data point into its own cluster. At each iteration step two closest clusters are merged and the representative for the new cluster is calculated. The merging criteria is based on either single or complete linkage rule depending on users preference. The merging of clusters is repeated until the user-defined number of final clusters is faced.

K-MEANS CLUSTERING: In k-means the cluster representatives are initialized by either user or random generation. At each iteration step the data to be clustered is gone through and each data vector (point) is assigned to the closest cluster after which the new representatives are determined. This is repeated until the convergence, that is, there are no changes in the representatives.

COMPETITIVE LEARNING: The applet is based on the "winner takes all" schema. The user can choose the learning rate and the representatives at the beginnig. At each iteration step one data vector x (point) is chosen and its closest representative is found. This representative w is updated by moving it into the direction of the data vector by using the equation below

w_new = w_old + learning_rate * [ x - w_old ]

After updating the representative, the learning rate is decreased.

MACBETHS SPECTRAL REFLECTANCES: The spectral reflectances of GretagMacbeth ColorChecker Color Rendition Chart at wavelength range from 380 nm to 780 nm will be shown to a user when clicking the patches. The spectra are measured by PR-705 spectroradiometer.

SPECTRUM TO RGB: Applet calculates the RGB coordinates for a reflectance spectrum drawn by the user. The used reference illuminant is D65.

HIGH- AND LOW-PASS FILTERING: Applet demonstrates the performance of high- and low-pass filtering of color spectrum. User can define both the incoming signal and the filter to be used.