Descriptors for Object Detection in Image Recognition

Abstract

The objects of interest can be described using various image information (e.g. shape, texture, colour). In the area of feature based detectors, the image features are the carries of this information. The goal is to successfully describe the object with a relatively small set of numbers; the large number of features slows down the training and detection phases and the methods for the reduction of feature vector must be used.

Many methods for extracting the image features that are able to describe the appearance of objects were presented, especially, the detectors that are based on the histograms of oriented gradients (HOG), Haar features, or local binary patterns (LBP) are dominant and they are considered as the state-of-the-art methods. Nevertheless, the classical features (e.g. HOG and Haar features) that are combined with the trainable classifiers (e.g. the support vector machine and neural network) require large training sets due to their high dimensionality. Unfortunately, the large training sets are difficult to acquire in many applications.

In this dissertation, we introduce an alternative and novel image descriptors that are based on the investigation of energy distribution (in the image) that describes the properties of objects. The energy distribution is encoded into a vector of features and the vector is then used as an input for the SVM classifier. Additional to this, we explore one another way how to encode the properties of objects. We propose method that is based on the fact that the properties of the image (especially the properties of the objects) can effectively be described by the distance function. Using the proposed approaches, the objects of interest can be successfully described with promising results and abilities, parameters, and improvements of the proposed descriptors are shown in this dissertation.