HMM based Viterbi paths for rejection correction in a convolutional neural network classifier
Abstract
This paper presents a rejection strategy for a convolutional neural network. The method is based on Viterbi paths generated by a context based 2D stochastic model for rejected image correction. The rejection strategy is an important issue in neural network theory. The challenge is to find rules, which determine if an image is correctly classified or not. Applying strong rules leads to the rejection of many well-recognized patterns whereas weak rules do not always involve a strong decrease of erroneous patterns. We propose to use in several ways the knowledge of an external stochastic model so called NSHP-HMM (Non Symmetric Half-Plane Hidden Markov Model) for re-evaluating the rejected patterns.