Improvements on Gabor Descriptor Retrieval for Patch Detection

Abstract

The localization of object parts in the component-based object detection is among the main tasks to solve. This paper presents several improvements of the proposed local image descriptor based on Gabor wavelets. Including these descriptors in the desired application is an ambitious challenge if we take into account the high number of parameters. Determining of parameters can be very hard because of their infinite definition range. Defining the filters is done in two stages: a theoretical consideration narrows the domain and the cardinality of parameters; this is followed by adequate experiments to select the most characteristic descriptor for a target image patch. The descriptor is created from a given number of 2D Gabor filters chosen by the GentleBoost learning algorithm. Comparing the proposed descriptor to those found in the state of the art, we can conclude that the selected filters are adaptable to any target object. In contrast to this, the majority of filter-based descriptors have fixed values for the parameters that do not allow to be ductile to the given object. Parameters fine-tuning allows the descriptor to be general, and discriminative at the same time. The effect of the following experiments has been analyzed during the investigation: elimination of redundancy between the weak classifiers, using the LoG interest points in the detection process. Finally, we propose an acceleration algorithm in order to deter- mine the response map faster. By means of the descriptor, the response map is created, which accurately localizes the target object part and can easily be integrated in almost all detection systems.