Search

Ancient Chinese character recognition is a challenging problem in the field of pattern recognition. It is difficult to collect all character classes during the training stage due to the numerous classes of ancient Chinese characters and the likelihood of discovering new characters over time. A solution to address this problem is continual learning. However, most continual learning methods are not well-suited for large-scale applications, making them insufficient for solving the problem of ancient Chinese character recognition. Although saving raw data for old classes is a good approach for continual learning to address large-scale problems, it is often infeasible due to the lack of data accessibility in reality. To solve these problems, we propose a large-scale continual learning framework based on the convolutional prototype network (CPN), which does not save raw data for old classes. In this paper, several basic strategies have been proposed for the initial training stage to enhance the feature extraction ability and robustness of the network, which can improve the performance of the model in continual learning. In addition, we propose two practical methods in varying feature space (parameters of feature extractor are changeable) and fixed feature space (parameters of feature extractor are fixed), which enable the model to carry out large-scale continual learning. The proposed method does not save the raw data of old classes and enables simultaneous classification of all existing classes without knowing the incremental batch number. Experiments on the CASIA-AHCDB dataset with 5000 character classes demonstrate the effectiveness and superiority of the proposed method. • We propose a prototype-based framework for large-scale continual learning. • The proposed strategies in the initial stage are good for continual learning. • We propose two continual learning methods in varying and fixed feature space. • The proposed methods achieve SOTA results on the CASIA-AHCDB dataset with 5000 classes. [ABSTRACT FROM AUTHOR]

Despite the success of convolutional neural network (CNN) in conventional closed-set recognition (CSR), it still lacks robustness for dealing with unknowns (those out of known classes) in open environment. To improve the robustness of CNN in open-set recognition (OSR) and meanwhile maintain its high accuracy in CSR, we propose an alternative deep framework called convolutional prototype network (CPN), which keeps CNN for representation learning but replaces the closed-world assumed softmax with an open-world oriented and human-like prototype model. To equip CPN with discriminative ability for classifying known samples, we design several discriminative losses for training. Moreover, to increase the robustness of CPN for unknowns, we interpret CPN from the perspective of generative model and further propose a generative loss, which is essentially maximizing the log-likelihood of known samples and serves as a latent regularization for discriminative learning. The combination of discriminative and generative losses makes CPN a hybrid model with advantages for both CSR and OSR. Under the designed losses, the CPN is trained end-to-end for learning the convolutional network and prototypes jointly. For application of CPN in OSR, we propose two rejection rules for detecting different types of unknowns. Experiments on several datasets demonstrate the efficiency and effectiveness of CPN for both CSR and OSR tasks. [ABSTRACT FROM AUTHOR]

Despite some interesting results from different research groups, a public database for Uyghur online handwriting recognition and a baseline study are not yet available for comparison purpose. In order to fill this void, we present a database of Uyghur online handwritten words and carry out the first benchmark experiments using it. This database contains 125,020 samples of 2030 words collected from 393 writers. According to Uyghur lexicon characteristics, two out-of-vocabulary datasets are especially provided for evaluation. We carry out some unconstrained handwritten word recognition experiments on the database using recurrent neural networks as base model. Recognition results are acquired using connectionist temporal classification without lexicon search and external language model. Concatenated and averaged bidirectional recurrent layers are compared for better generalization. Based on Uyghur unicode representation, we are interested in comparing the models using different alphabets, based both on character types and character forms. To improve generalization, we propose 1D convolutional model which implements 1D convolutional layers for sequence feature extraction. In our experiments, the proposed 1D convolutional model and its variations surpassed the base recurrent layered model on the out-of-vocabulary words by clear margin. 83.23% CAR (character accurate rate) was resulted when out-of-vocabulary samples are used for testing. The highest recognition rate is as high as 94.95% CAR when the test set shares the same lexicon to the training set. The experiments in this paper can be the baseline references for the future study using this database. [ABSTRACT FROM AUTHOR]

• A unified theoretical framework for long-tailed recognition is established. • Corresponding mitigation solutions for prior gap and representation gap are proposed. • Theoretically analyzing the existing methods and the proposed methods in terms of the impact on two gaps. • The proposed methods yield superior performance on five long-tailed benchmarks. Most deep learning models are elaborately designed for balanced datasets, and thus they inevitably suffer performance degradation in practical long-tailed recognition tasks, especially to the minority classes. There are two crucial issues in learning from imbalanced datasets: skew decision boundary and unrepresentative feature space. In this work, we establish a theoretical framework to analyze the sources of these two issues from Bayesian perspective, and find that they are closely related to the prior gap and the representation gap, respectively. Under this framework, we show that existing long-tailed recognition methods manage to remove either the prior gap or the presentation gap. Different from these methods, we propose to simultaneously remove the two gaps to achieve more accurate long-tailed recognition. Specifically, we propose the prior calibration strategy to remove the prior gap and introduce three strategies (representative feature extraction, optimization strategy adjustment and effective sample modeling) to mitigate the representation gap. Extensive experiments on five benchmark datasets validate the superiority of our method against the state-of-the-art competitors. [ABSTRACT FROM AUTHOR]

In this paper, we propose an adaptive spatial pooling method for enhancing the discriminability of feature representation for image classification. The core idea is to adopt a spatial distribution matrix to define how the image patches are pooled together. By formulating the pooling distribution learning and classifier training jointly, our method can extract multiple spatial layouts of arbitrary shapes rather than regular rectangular regions. By proper mathematical transformation, the distributions can be learned via a boosting-like algorithm, which improves the efficiency of learning especially for large distribution matrices. Further, our method allows category-specific pooling operations to take advantage of the different spatial layouts of different categories. Experimental results on three benchmark datasets UIUC-Sports, 21-Land-Use and Scene 15 demonstrate the effectiveness of our method. [ABSTRACT FROM AUTHOR]

Abstract: This paper presents a new topic of automatic recognition of bank note serial numbers, which will not only facilitate the prevention of forgery crimes, but also have a positive impact on the economy. Among all the different currencies, we focus on the study of RMB (renminbi bank note, the paper currency used in China) serial numbers. For evaluation, a new database NUST-RMB2013 has been collected from scanned RMB images, which contains the serial numbers of 35 categories with 17,262 training samples and 7000 testing samples in total. We comprehensively implement and compare two classic and one newly merged feature extraction methods (namely gradient direction feature, Gabor feature, and CNN trainable feature), four different types of well-known classifiers (SVM, LDF, MQDF, and CNN), and five multiple classifier combination strategies (including a specially designed novel cascade method). To further improve the recognition accuracy, the enhancements of three different kinds of distortions have been tested. Since high reliability is more important than accuracy in financial applications, we introduce three rejection schemes of first rank measurement (FRM), first two ranks measurement (FTRM) and linear discriminant analysis based measurement (LDAM). All the classifiers and classifier combination schemes are combined with different rejection criteria. A novel cascade rejection measurement achieves 100% reliability with less rejection rate compared with the existing methods. Experimental results show that MQDF reaches the accuracy of 99.59% using the gradient direction feature trained with gray level normalized data; the cascade classifier combination achieves the best performance of 99.67%. The distortions have been proved to be very helpful because the performances of CNNs boost at least 0.5% by training with transformed samples. With the cascade rejection method, 100% reliability has been obtained by rejecting 1.01% test samples. [Copyright &y& Elsevier]