Detecção de objetos em imagens por meio da combinação de descritores locais e classificadores

Abstract

The detection of objects in images remains one of the biggest challenges in computer vision area because the objects that are contained in images may be under the most varied perspectives and changes of scale and rotation, which makes more complex the goal to detect them. This task has applications in many different contexts, ranging from medical diagnosis and the business area until regards to public safety. In order to seek a solution to the problem of detection of generic objects, a module that is based on the detection via classification was developed. The implemented module uses local descriptors to compute the features invariant to transformations of images, the bag-of-Keypoints model based by concepts of area information retrieval, to perform the pro-cessing of data extracted from the images to be compatible as entrance to the evalu-ated inductors. Were evaluated, combinations of two local descriptors (SIFT and SURF) with five supervised learning approaches, the algorithms: Multilayer Percep-tron, FURIA, Random Forest, Support Vector Machines and the k-nearest neigbor. From the module construction and analysis of experimental results and real scenarios, it was found that the models generated by the built module have high accuracy rates in experimental settings and promising results in real scenario.