Explorando limitações das redes convolucionais na seleção de testes binários

Abstract

Feature extraction is a fundamental step in Computer Vision systems, which consists in transforming the raw pixels data into new representations robust to illumination, rotation and scale variations. These new representations are known as feature vectors and can be generated by different methodologies depending on the application requirements. For many years algorithms based on gradient orientation histograms, such as SIFT, HOG and SURF, were the most efficient for low level feature extraction but they have a high computational cost due mainly to the use of convolutions. To avoid this problem some algorithms have been created to generate binary vectors, such as BRIEF, ORB, BRISK and FREAK. They propose different spatial distributions of binary tests and the main goal is to keep a low cost to be executed on small and large computers. Recent advances in feature extraction have been obtained using deep learning and the results of Convolutional Neural Networks (CNN) outperformed the handcrafted descriptors. Different CNN architectures were created by researchers and some of them are projected to generate binary vectors, such as LIFT and DeepBit. Their strategies are focused in binarizing the Real values of output layer, which maintains a high computational cost. This work investigates the problem of selecting binary tests in order to discover different spatial distributions that allows the creation of a new binary descriptor. Some limiting properties for searching solutions with CNNs are presented, which appears when the gradient is calculated based on the local pixels neighborhood. Experiments were conducted using a Siamese Network trained with corresponding and non-corresponding patch pairs, whose results show the existence of local minimum and another problem defined in this dissertation as Incorrect Components of Gradient.