Deformations and Discriminative Models for Image Recognition

In this work we present approaches to incorporate domain-knowledge into discriminative classifiers. In particular, we investigate the incorporation of the image distortion model into log-linear models and support vector machines. Discriminative models are a well-known technique in many fields of machine learning and pattern recognition and the image distortion model has recently been shown to be a very effective means of modelling variability in images of handwritten characters, however, so far it was impossible to fuse the advantages of these two different approaches. In order to incorporate the IDM into the log-linear model, we re-investigate the probabilistic formulation of the IDM, and the relationships between Gaussian models and log-linear models, which allows for a direct integration of the IDM into the log-linear framework, where an alignment of an image to a prototype is considered a hidden variable. Alternating optimisation techniques allow for effective training of log-linear models with hidden variables. Furthermore,we have investigated how to combine SVMs with the IDM. The performance of the different models is examined on two standard tasks, the USPS tasks, which, due to its small size, allows for extensive testing and tuning of models and the MNIST database, which is considered a standard benchmark in OCR applications. The results obtained compare favourably well to other, comparable models. In particular, we report the best error rate using a single density model on the USPS task.