Deep Boltzmann Machines based vehicle recognition

We address the problem of image-based vehicle recognition. This problem is formulated as a manifold-feature learning algorithm such that the expected object distribution is evaluated based on the appearance, shape and contours that take into account variational poses and occlusions. We propose to learn manifold hierarchical features where the high-level discriminative features are obtained by capturing correlations among the learned low-level generic features, via a deep learning model, called Deep Boltzmann Machines (DBM), a powerful hierarchical generative model for feature learning. DBM has been proved to be able to explore underlying manifold knowledge of given objects by appropriately describing highly nonlinear complicated functions with novel training process. Moreover, we also present a novel application of deep learning to feature representations, rather than provide the model input with raw image pixels, we instead utilize the human-engineered descriptors such as Log-Gabor, HoG and Gist, as the source data of deep learning architecture. Finally, we evaluate our deep model on PASCAL VOC2012 benchmark dataset, extensive experiments show that our DBM-based model enables to learn useful hierarchical feature representations from few training samples. Specifically, by feeding the DBM-based deep network with prior-described features such as Log-Gabor, HoG and Gist, we get a significant improvement on performance in comparison with the state-of-art results and the best performance is achieved by using a fusion form of the three descriptors as source material for learning deep model. Furthermore, experiments with prior-extracted features as input layer demonstrate that the deep learning networks have provided a perspective of potential ability to learn critical representation of given data, and have presented favorable vehicle recognition performance.