Recently, there has been an increasing interest in developing diffusion-based text-to-image generative models capable of generating coherent and well-formed visual text. In this paper, we propose a novel and efficient approach called GlyphControl to address this task. Unlike existing methods that rely on character-aware text encoders like ByT5 and require retraining of text-to-image models, our approach leverages additional glyph conditional information to enhance the performance of the off-the-shelf Stable-Diffusion model in generating accurate visual text. By incorporating glyph instructions, users can customize the content, location, and size of the generated text according to their specific requirements. To facilitate further research in visual text generation, we construct a training benchmark dataset called LAION-Glyph. We evaluate the effectiveness of our approach by measuring OCR-based metrics, CLIP score, and FID of the generated visual text. Our empirical evaluations demonstrate that GlyphControl outperforms the recent DeepFloyd IF approach in terms of OCR accuracy, CLIP score, and FID, highlighting the efficacy of our method.

Recently Adam has become a popular stochastic optimization method in deep learning area. To parallelize Adam in a distributed system, synchronous stochastic gradient (SSG) technique is widely used, which is inefficient due to heavy communication cost. In this paper, we attempt to parallelize Adam with blockwise model-update filtering (BMUF) instead. BMUF synchronizes model-update periodically and introduces a block momentum to improve performance. We propose a novel way to modify the estimated moment buffers of Adam and figure out a simple yet effective trick for hyper-parameter setting under BMUF framework. Experimental results on large scale English optical character recognition (OCR) task and large vocabulary continuous speech recognition (LVCSR) task show that BMUF-Adam achieves almost a linear speedup without recognition accuracy degradation and outperforms SSG-based method in terms of speedup, scalability and recognition accuracy.

By leveraging large amounts of training data and deep learning technologies, performances of modern handwritten optical character recognition (OCR) systems have been greatly improved. However, collecting and labeling massive handwriting images are both time-consuming and expensive. In this paper, we propose to augment handwritten OCR training with online handwriting samples. To achieve this goal, we propose a style-conditioned generative adversarial network (SC-GAN) with a novel training data pair generation strategy. Then this network is used to transfer the styles of real handwriting images to skeleton images extracted from online handwriting samples to generate photo-realistic text line images. Experimental results on a large scale handwritten OCR task show that the recognition accuracy of our handwritten OCR system is improved by using the augmented synthetic training data.

Knowledge distillation (KD) has been widely used to improve the performance of a simpler student model by imitating the outputs or intermediate representations of a more complex teacher model. The most commonly used KD technique is to minimize a Kullback-Leibler divergence between the output distributions of the teacher and student models. When it is applied to compressing acoustic models trained with a connectionist temporal classification (CTC) criterion, an assumption is made that the teacher and student share the same frame-level feature-transcription alignment. However, frame-level alignments learned by teachers can be inaccurate and unstable due to the lack of fine-grained frame-level guidance during CTC training. Forcing student to learn inaccurate alignments will lead to limited performance improvements. In this article, we investigate building powerful teacher models with more accurate and stable feature-transcription alignments. We achieve this goal by using a novel alignment-consistent ensemble (ACE) technique, where all models within an ensemble are jointly trained along with a regularization term to encourage consistent and stable alignments. With well-trained deep bidirectional LSTM (DBLSTM) ACE as a teacher, we can directly use the traditional frame-wise KD method to train DBLSTM students. When applying KD to transfer knowledge from a DBLSTM ACE to a deep unidirectional LSTM (DLSTM) student, a simple yet effective target delay technique is proposed to handle the alignment difference between bidirectional and unidirectional models. Experimental results on Switchboard-I speech recognition task show that, with DBLSTM ACE as a teacher, the simple frame-wise KD method can achieve competitive or better performance than other complex KD methods on DBLSTM students. When applying KD to build DLSTM students from DBLSTM teachers, our proposed target delay technique can achieve relative word error rate reductions of 14.2%$\sim$14.8% compared with the models trained from scratch, which outperforms other carefully-designed KD methods.

Character models based on convolutional neural network (CNN) and deep bidirectional long short-term memory (DBLSTM) have achieved high recognition accuracy on various handwriting recognition (HWR) and OCR tasks. To deploy CNN-DBLSTM models in products, it is necessary to reduce the footprint and runtime latency as much as possible. In this paper, we use a teacher-student learning approach to achieve this goal, where a new objective function is proposed to match the extracted CNN feature sequences of the teacher and student models under the guidance of the succeeding LSTM layer. Experimental results on large scale English HWR and OCR tasks show that the learned small student model can achieve about 14.6x footprint reduction and 9.6x speedup without recognition accuracy degradation against the big teacher model.

Recently, character model based on integrated convolutional neural network (CNN) and deep bidirectional long short-term memory (DBLSTM) has been demonstrated to be effective for online handwritten Chinese text recognition (HCTR). However, the reported CNN-DBLSTM topologies are too complex to be practically useful. In this paper, we propose a compact CNN-DBLSTM which has small footprint and low computation cost yet be able to accommodate multiple receptive fields for CNN-based feature extraction. By using the training set of a popular benchmark database, namely CASIA-OLHWDB, we trained a compact CNN-DBLSTM by a connectionist temporal classification (CTC) criterion with a multi-step training strategy. Combined this character model with a character trigram language model, our online HCTR system with a WFSTbased decoder has achieved state-of-the-art performance on both CASIA and ICDAR-2013 Chinese handwriting recognition competition test sets.

Recently, character model based on integrated convolutional neural network (CNN) and deep bidirectional long short-term memory (DBLSTM) has achieved excellent performance for offline handwriting recognition (HWR). To deploy CNN-DBLSTM model in products, it is necessary to reduce the footprint and runtime latency as much as possible. In this paper, we study two methods to compress the CNN part: (1) Use Tucker decomposition to decompose pre-trained weights with low-rank approximation, followed by fine-tuning; (2) Use grouped convolution to construct sparse connections in channel domain. Experiments have been conducted on a large-scale offline English HWR task to compare the effectiveness of the above two techniques. Our results show that using Tucker decomposition alone offers a good solution to building a compact CNN-DBLSTM model which can reduce significantly both the footprint and latency yet without degrading recognition accuracy.

We study two sequence discriminative training criteria, i.e., Lattice-Free Maximum Mutual Information (LFMMI) and Connectionist Temporal Classification (CTC), for end-to-end training of Deep Bidirectional Long Short-Term Memory (DBLSTM) based character models of two offline English handwriting recognition systems with an input feature vector sequence extracted by Principal Component Analysis (PCA) and Convolutional Neural Network (CNN), respectively. We observe that refining CTC-trained PCA-DBLSTM model with an interpolated CTC and LFMMI objective function ("CTC+LFMMI") for several additional iterations achieves a relative Word Error Rate (WER) reduction of 24.6% and 13.9% on the public IAM test set and an in-house E2E test set, respectively. For a much better CTC-trained CNN-DBLSTM system, the proposed "CTC+LFMMI" method achieves a relative WER reduction of 19.6% and 8.3% on the above two test sets, respectively.