Machine recognition of Chinese handwriting finds many applications. For example, mail address recognition leads to automatic mail sorting which saves a lot of human labor. Bank check reading, tax form processing, book and handwritten notes transcription transfer documents into digital format which is convenient for administration, search and transmission. Handwritten Chinese character recognition (HCCR), which is an integral part of handwritten Chinese text recognition, has drawn much attention from the community. The difficulty of HCCR lies in the large number of character classes, the presence of many confusing character pairs, and the variability of writing styles. Due to these difficulties, the performance of HCCR is still not satisfactory. In this thesis, based on traditional recognition methods, we boost the performance of HCCR with two methods -- accelerating classifier training and large-scale feature learning. The proposed methods are summarized as follows. 1. Accelerating classifier training using graphics processing units (GPU). For enhancing the generalization performance of classifiers, training set expansion is an effective strategy usually adopted. However, increasing training set incurs long training time for classifier training, especially for those methods based on discriminative learning. In this thesis, we propose to accelerate the training of discriminative feature extraction (DFE) and discriminative learning quadratic discriminant function (DLQDF) classifier by parallelizing the computation using GPU. Thirty times and ten times speedup was achieved, respectively. 2. Improving recognition accuracy through large-scale feature learning. For enhancing the discrimination ability of features, we increase the feature dimensionality by using statistical and spatial correlation of original gradient direction histogram features. This generates tens of thousands of quadratic features. A low-dimensional subspace is learned from the quadratic features and original gradient features by discriminative learning. For the integration of quadratic information and discriminative learning, the resultant subspace features are quadratic as well as discriminative. To further improve the generalization capability of learned features as well as classifiers, we expand the training set with synthesized samples. In experiments of HCCR, with the proposed feature learning method and DLQDF classifier, we achieved recognition performance comparable to deep conv...
修改评论