Reconocimiento robusto de lugares mediante redes neuronales convolucionales

In this work, we propose using Convolutional Neural Network for place visual recognition. The work focuses on the identification and automatic extraction of interest regions from a query image. These regions are used to build an image encoding through a vector of locally aggregated descriptors, which is turn used for image recovery. Unlike other methods, where the entire image is used to create the encoding, our approach only uses the most important image interest regions. This provides better invariance to changes at extreme view points of view, lighting and occlusions. Another contribution of the work consists in the integration of a totally convolutional spatial transformer according to the convolutional neural network architecture. This transformer is used for normalizing these interest regions, which allows achieving a greater robustness during coding. A loss function is also proposed that is used to train the artificial neural network to identify automatically regions. To measure the efficiency of the proposed model, a variety of experiments were carried out with challenging data sets. The reported results show that the proposed method produces superior results than other state of the art methods.