A novel feature selection method using generalized inverted Dirichlet-based HMMs for image categorization

Hidden Markov Models (HMMs) have consistently been a powerful tool for performing numerous challenging machine learning tasks such as automatic recognition. The latter perceives all objects of the universe through information carried by their characteristics or features. However, not all available data is always valuable for distinguishing between the different objects, scenes, scenarios; referring analogically to states. More often than not, automatic recognition is accompanied by a feature selection to reduce the number of collected features to a relevant subset. Although sparse, the majority of literature resources available on feature selection for HMMs, presuppose either a single Gaussian or employ a Gaussian mixture model (GMM) as emission distribution. The proposed method builds upon the feature saliency model introduced by Adams, Cogill, and Beling (in IEEE Access 4:1642–1657), and is adjusted to handle complex multidimensional data by using as a novel experiment, GID (Generalized Inverted Dirichlet) mixture models) as emission probabilities. We make use of an Expectation-Maximization (EM) algorithm (Dempster et al. in J R Stat Soc 39(1):1–22) to compute maximum a posteriori (MAP) [Gauvain and Lee in IEEE Transact Speech Audio Process 2(2):291–298] estimates for model parameters. The complete inference and parameter estimation of our GID-FSHMM (GID Feature Selection-based HMM) are detailed in this work. Automatic recognition applications such as facial expression recognition and scenes categorization demonstrate comparable to higher performance compared to the extensively used Gaussian mixture-based HMM (GHMM), the Dirichlet-based (DHMM) and the inverted Dirichlet-based HMM (IDHMM) without feature selection and also when the latter is embedded in all of the aforementioned models.