Distance Based k-NN Classification of Gabor Jet Local Descriptors

Abstract

Abstract In the domain of artificial intelligence object detection has become a popular area over the past several years. The creation of an automatic detection system plays an important role in several domains of interest, such as: bioinformatics, traffic supervision, access control, identification and authentication systems and industry using intelligent robots etc. Creating such a detection system is a challenge for every researcher in this domain. The main difficulty comes from the extreme diversity in which all objects appear. They have a large variety of appearance, aspect, form, dimension, color, position, rotation angle, illumination, shadow or occlusion. In this approach we analyzed part-based object detection systems. These can be generally separated in three main phases: detection of interest points, local descriptor and the object model. This paper proposes a new local descriptor, for the second phase and compares its detection performance with several classification algorithms. The developed patch descriptor is based on two-dimensional Gabor wavelets. The Gabor filters are Gaussian modulated sinusoidal waves, which describe the neighborhood of a given image pixel in two-dimensional space. It is defined by 9 degrees of freedom. Each of these parameters can have an infinite definition domain. The goal is to reduce the infinite number of possible values and to determine the most adequate filters for a given object. The definition domain of the nine parameters is narrowed by some theoretical considerations and by the dimension of the image patch analyzed. According to our experiments, we have deduced set of Gabor filter responses which characterize the region of interest in a given image. The goal of this approach is to find the most characteristic Gabor filters for the object of interest. After defining more thousand such filters, with a selection algorithm, we determine the most discriminative n filters based on the training set of images and the total number of defined Gabor filter descriptors. Only the first best n classifiers are going to play a role in the classification of the image patch. This paper compares the k-NN decision to other learning methods as the Gentle Boost algorithm, the SVM classification, which we have used in our previous works. The choice of Gabor wavelets for object detection is well-founded, because it has been physiologically demonstrated that the human visual cortex system works similarly, in other words, it can be modeled by Gabor filter decomposition [5] , [6] . In our previous works we have defined a novel Gabor-filter based patch descriptors for object detection. In this approach we classify them with different classification methods. We have obtained high classification performance with an easy, computationally simple algorithm as k-NN Nearest Neighbors method, which reduces the training process. Our contribution is finding the most adequate Gabor filters parameters considering a given object or object part. After defining a novel patch descriptor based on these responses, we compare several classification methods, in order to obtain as good detection performances as possible. The main contribution of this paper is to apply a simple, but efficient classification method as the k-NN, in order to reduce the computations compared to previously used classification algorithms [1] , [2] , [3] , [4] .