Sufficient Statistics Feature Mapping over Deep Boltzmann Machine for Detection

As an important and fundamental methodology in the fields of pattern recognition and image processing, learning middle level feature has attracted increasing interest during the recent years, where generative feature mapping has shown highly completive performance in diverse applications. In this paper, a middle level feature representation is proposed based on Deep Boltzmann Machine (DBM) and sufficient statistics (SS) feature mapping for detection. In the approach, DBM is employed to model data distribution and the hidden information inferred by DBM together with other informative variables are then exploited by SS to form the middle level features. The features, learnt from data, can be fed to standard classifiers for classification. In order to evaluate the performance of our method, we apply our feature mapping method to two challenging tasks: (1) contour detection through distinguishing border and non-border pixels, (2) sales pipeline prediction, which predicts the winning propensity of the ongoing sales opportunity in the pipeline. In comparison with other leading methods in the literature on the Berkeley Segmentation Dataset and Sales Pipeline Database (SPDB), our proposed algorithm performs favorably against state-of-the-art methods in terms of effectiveness and efficiency.