Classifier Introducing Transition Likelihood Model Based on Quantization Residual

Binary codes that are binarizations of features represented by real numbers have recently been used in the object recognition field, in order to achieve reduced memory and robustness with respect to noise. However, binarizing features represented by real numbers has a problem in that a great deal of the information within the features drops out. That is why we focus on quantization residual, which is information that drops out when features are binarized. With this study, we introduce a transition likelihood model into classifiers, in order to take into consideration the possibility that a binary code which has been observed from an image will transition to another binary code. This enables classifications that consider transitions to the desired binary code, even if the observed binary code differs from the actually desired binary code for some reason. From the results of experiments, we confirmed that the proposed method enables an increase in detection performance while maintaining the same levels of memory and computing costs as those for previous methods of binarizing features.