Handwritten Character Recognition by Alternately Trained Relaxation Convolutional Neural Network

Abstract

Deep learning methods have recently achieved impressive performance in the area of visual recognition and speech recognition. In this paper, we propose a handwriting recognition method based on relaxation convolutional neural network (R-CNN) and alternately trained relaxation convolutional neural network (ATR-CNN). Previous methods regularize CNN at full-connected layer or spatial-pooling layer, however, we focus on convolutional layer. The relaxation convolution layer adopted in our R-CNN, unlike traditional convolutional layer, does not require neurons within a feature map to share the same convolutional kernel, endowing the neural network with more expressive power. As relaxation convolution sharply increase the total number of parameters, we adopt alternate training in ATR-CNN to regularize the neural network during training procedure. Our previous C-NN took the 1st place in ICDAR'13 Chinese Handwriting Character Recognition Competition, while our latest ATR-CNN outperforms our previous one and achieves the state-of-the-art accuracy with an error rate of 3.94%, further narrowing the gap between machine and human observers (3.87%).

Discriminative quadratic feature learning for handwritten Chinese character recognition

“…We can see that the best accuracy of DLQDF is 94.44%, which is comparable to the deep CNN's 94.47% announced in [7], but still inferior to relaxation CNN's 95.04% [8]. In [8], weights sharing restriction of convolutional layers is relaxed to enhance the expressive capability, and at the same time, the number of parameters is largely increased. Unlike these deep CNNs, no extra synthesized samples are used for training in this experiment.…”

“…Comparing Table 18 with Table 4, by combining DQFL and training set expansion, the test accuracies of NPC, MQDF and DLQDF are improved by 5.94%, 2.53% and 2.62%, respectively; and meanwhile, the accuracy gap between NPC and DLQDF is reduced from 4.08% to 0.76%, which demonstrates that quadratic features are more beneficial for linear classifiers than for quadratic classifiers. As shown in Table 19, on the ICDAR-2013 test set, the best accuracy of DLQDF reaches 94.92%, which is close to relaxation CNN's 95.04% [8], but significantly inferior to CNN committee's 96.06% [8]. When considering test speed, our system is much faster than CNN as analyzed in next subsection.…”

“…Usually, a large training set produced by affine transform or elastic distortion is used to improve its generalization ability. In a recent HCCR competition [5], deep CNN committee reported test accuracies comparable to human observers [7,5,8]. However, the computation burden of deep CNNs is much higher than that of traditional methods due to multiple layers of convolution.…”

“…Deep CNNs, which are comprised of multiple layers of convolution and pooling for automatic feature learning and fully connected layers for classification, have outperformed traditional methods by a large margin [7,5,8]. The deep hierarchical structure of CNNs enables it to learn high-level features relevant for recognition.…”

“…Traditional HCCR methods mostly use hand-crafted features, and in recent years, some works based on feature learning using deep neural networks, especially, deep convolutional neural networks (CNNs), have reported superior performance [6][7][8]. Traditional methods recognize a sample in a pipeline of shape normalization, feature extraction, dimensionality reduction and classification.…”