Morphological analysis for automatized visual inspection using reduced HOG

This paper presents a methodology for the development of object detection and classification systems in which morphology is the major discriminating feature. This methodology is based on a common descriptor known as Histogram of Oriented Gradients, HOG, and it uses a support vector machine for classification. Regular HOG is a high dimensionality feature vector and its computation is the time bottleneck for those applications based on it. We propose a systematic dimensionality reduction of HOG features by means of identifying those descriptor blocks that have no or low discriminatory power, and thus are not necessary to compute. In our method, this non-contributing blocks are selected during training, by computing a reliability index for the overall system while removing one block at a time. Experiments and performance evaluation were carried out on two standardized databases commonly used in pedestrian detection and with a locally generated database for the quality assurance of plastic bottles. Qualitative results on pedestrian detection from a moving vehicle are also shown.