Incremental Learning Vector Quantization for Character Recognition with Local Style Consistency

Abstract

Incremental learning is a way relevant to human learning that utilizes samples in online sequence. In the paper, we propose an incremental learning method called Incremental Adaptive Learning Vector Quantization (IALVQ) which aims at classifying characters appearing in an online sequence with style consistency in local time periods. Such local consistency is present commonly in document images, in that the characters in a paragraph or text line are printed in the same font or written by the same person. Our IALVQ method updates the prototypes (parameters of classifier) incrementally to adapt to drifted concepts globally while utilize the style consistency locally. For style adaptation, a style transfer mapping (STM) matrix is calculated on a batch of samples of assumed same style. The STM matrix can be used both in training for prototypes updating and in testing for labels prediction. We consider supervised incremental learning and active incremental learning. In the latter way, class labels are attached only to samples that are assigned low confidence by the classifier. In our experiments on handwritten digits in the NIST Special Database 19, we evaluated the classification performance of IALVQ in two scenarios, interleaved test-then-train and style-specific classification. The results show that utilizing local style consistency can improve the accuracies of both two test scenarios, and for both supervised and active incremental learning modes.