Visual attention models for scene text recognition
In this paper we propose an approach to lexicon-free recognition of text in scene images. Our approach relies on a LSTM-based soft visual attention model learned from convolutional features. A set of feature vectors are derived from an intermediate convolutional layer corresponding to different areas of the image. This permits encoding of spatial information into the image representation. In this way, the framework is able to learn how to selectively focus on different parts of the image. At every time step the recognizer emits one character using a weighted combination of the convolutional feature vectors according to the learned attention model. Training can be done end-to-end using only word level annotations. In addition, we show that modifying the beam search algorithm by integrating an explicit language model leads to significantly better recognition results. We validate the performance of our approach on standard SVT and ICDAR’03 scene text datasets, showing state-of-the-art performance in unconstrained text recognition.
The increasing ability to capture images in any condition and situation poses many challenges and opportunities for extracting visual information from images. One such challenge is the detection and recognition of text “in the wild”. Text in natural images is high level semantic information that can aid automatic image understanding and retrieval.
Text can also play a role in a number of applications such as automatic translation, assisting the visually impaired, and robot navigation .However, robust reading of text in uncontrolled environments is challenging due to a multitude of factors such as difficult acquisition conditions, low resolution, font variability, complex backgrounds, different lighting conditions, blur, etc.
Recognizing text in outdoor scene images is very different than text recognition in document images, where the problem is well studied and many commercially successful Optical Character Recognition (OCR) systems exist. OCR techniques used in document images do not generalize to recognition of scene text because they are tuned for document images having black text on white backgrounds and are mostly scanned using flatbed scanner with uniform lighting and exposure.
The problem of end-to-end scene text recognition is usually divided in two different tasks: word detection and word recognition. The goal of the word detection stage is to generate bounding boxes around potential words in the images. Subsequently, the words in these bounding boxes are recognized in the word recognition stage. This paper is focused on this second stage, word recognition.
Many existing scene text recognition methods rely on a predefined lexicon or dictionary [1, 8, 13, 17, 19, 20, 22], which greatly improves accuracy by constraining output. Although the use of a predefined lexicon restricts the set of possible words to be recognized, excellent results have been reported recently with dictionaries as large as 90,000 words . Lexicon-based recognition works because the lexicon acts as language prior or a recognition context. However, in many cases such contextual information is either unavailable or hard to determine. For instance, in case of images containing proper nouns like names of brands or products, street names, etc. To deal with such problems unconstrained text recognition methods have also been proposed [4, 10, 14] (here by unconstrained we mean that any word can be recognized). In this work we focus on unconstrained text recognition.
A class of methods for scene text recognition [4, 22] use an over-segmentation to generate multiple hypotheses of character locations, and then follow with a supervised character classifier. Word recognition is the result of finding the best sequence of character hypotheses according to classifier scores and a set of spatial or lexical constraints. These methods fit the unconstrained scenario quite naturally, but can be limited by the performance of character segmentation which is a challenging step due to the nature of scene text. Alternatively, another set of methods use a fixed-length representation to represent word images holistically and avoid the segmentation step. Then this representation is used either to classify words images among 1 of predefined word classes  or rank words from a fixed lexicon [1, 19] based on some sort of distance criteria. These approaches are mostly linked to lexicon-based word recognition although in some cases they have also been applied to unconstrained text recognition by inferring characters and -grams from the fixed-length representation . However, this increases the training complexity and requires incremental training and heuristic gradient rescaling.
In contrast to the above strategies our approach neither recognizes individual characters in the word image nor uses any holistic representation to recognize the word. It rather uses a LSTM-based visual attention model (based on ) to focus attention on relevant parts of the image at every step and infer a character present in the image. Thus, the system is able to recognize out-of-vocabulary words, although it does not need explicit character segmentation or recognition. The visual attention model can be trained using only word bounding boxes and does not need explicit character bounding boxes at training time. As the model relies on recurrent neural networks, it also learns an implicit language model from the data, which is crucial in cases such as those shown in . However, the model also has the flexibility to integrate an explicit language model, which it is shown in the experiments to improve accuracy. Additionally, the output can be constrained to a fixed lexicon (when available) to work in a dictionary-based setting. An overview of the proposed recognition framework is illustrated in figure 1.
The rest of the paper is organized as follows. In Section 2 we analyze the works most related to our proposed approach. Then, we present our attention-based recognition approach in Sections 3 and 4. In Section 5 we experimentally validate the model on a variety of standard and public benchmark datasets. We conclude in Section 6 with a summary of our contributions and a discussion of future research directions.
2 Related Work
In this section we briefly review work from the literature on robust scene text recognition and attention-based recognition models most related to our approach.
Dictionary-based scene text recognition. Traditionally, scene text recognition systems use character recognizers in a sequential way by localizing characters using a sliding window [13, 17, 22] and then grouping responses by arranging the character windows from left to right as words. A variety of techniques have been used to classify character bounding boxes, including random ferns , integer programming  and Convolutional Neural Networks (CNNs) . These methods often use the lexical constrains imposed by a fixed lexicon while grouping the character hypotheses into words.
In contrast to sequential character recognizer models, holistic fixed-length representations have been proposed in [1, 8, 19]. These works advocate the use of a joint embedding space between images and words. In addition, Gordo et al. in  make use of supervised mid-level feature learned using character bounding boxes to further improve the image-text embedding. In contrast, Yao et al. in , used mid-level features which can be learned from the data in a unsupervised way.
More recently, with the success of deep-CNN features in computer vision, convolutional features have also been applied to scene text recognition. The first such attempt was made by Jaderberg et al. in , where a sliding window over CNN features is used for robust scene text recognition using a fixed lexicon. Later, the same authors also proposed a fixed-length representation  using convolutional features trained of a synthetic dataset of 9 million images 
Unconstrained scene text recognition. Though most of the works in scene text recognition focus on fixed-lexicon recognition, a few attempts at unconstrained text recognition have also been made.
Biassco et al. in  rely on sequential character classifiers. They use a massive number of annotated character bounding boxes to learn character classifiers. Binarization and sliding window methods are used to generate character proposals followed by a text/background classifier. Finally, character probabilities given by character classifiers are used in a beam search to recognize words. They also integrate a static character -gram language model in every step of the beam search to incorporate an underlying language model.
Though CNN models have achieved great success in lexicon-based text recognition, word recognition in unconstrained scenarios requires modeling the underlying character-level language model. Jaderberg et al. in  proposed to use two separate CNNs, one modeling character unigram sequences and another -gram language statistics. They additionally use a Conditional Random Field to model the interdependence of characters (-grams). However, this significantly increases computational complexity. In addition, to detect the presence of character -grams in word images as neural activations, character -grams are used as output nodes, leading to a huge (10k output units for =4) output layer.
To model inter-dependencies between characters, the authors of  used a recursive CNN and variants of Recurrent Neural Networks on top of CNN features.
Visual attention models for recognition. Recently visual attention models have gained a lot of attention and have been used for machine translation  and image captioning . In this last work the attention model is combined with an LSTM on top of CNN features. The LSTM outputs one word at every step focusing on a specific part of the image driven by the attention model. Two models of attention, hard and soft attention are proposed. In our work, we mainly follow the soft attention model, adapted to the particular case of text recognition. Attention models appear to have the potential ability to overcome some of the limitations of existing text recognition methods. They can leverage a fixed-length representation, but at the same time, they are able to guide recognition to relevant parts of the image, performing in this way a kind of implicit character segmentation.
Recently, a soft attention model has also been proposed for text recognition in the wild . The main differences with our work are the following. Firstly, the success of  can be largely attributed to the use of Recursive Neural Network (RNN) features. They rely on the RNN features to model the dependencies between characters. Instead we use traditional CNN features and it is the visual attention model who learns to selectively attend to parts of the image and the dependencies between them. Secondly, Lee et al.  used the features from fully connected layer, while we use features from an earlier convolutional layer, thus preserving the local spatial characteristics of the image and reducing the model complexity. This also allows the model to focus on a subset of features corresponding to certain area of the image and learn the underlying inter-dependencies. Thirdly, we used LSTM instead of RNN which has been shown to learn long term dependencies better than traditional RNNs.
Our contributions with respect to the state-of-the-art. In summary the contributions of our work are:
We introduce a LSTM-based visual attention model for unconstrained scene text recognition. This model is able to selectively attend to specific parts of word images, allowing it to model inter-character dependencies as needed and thus to implicitly model the underlying language.
We show that weak explicit language models (in the form of prefix probabilities) can significantly boost the final recognition result without having to resort to a fixed lexicon. For that, We modify the beam search to take into account the language model. Additionaly, the beam search can also incorporate a lexicon whenever it is available.
We experimentally validate that our approach with weak language modeling outperforms the state-of-the-art in unconstrained scene text recognition and performs comparably to lexicon-based approaches with a model complexity lower than similar approaches.
3 Visual attention for scene text recognition
Our recognition approach is based on an encoder-decoder framework for sequence to sequence learning. An overall scheme of the framework is illustrated in figure 2. The encoder takes an image of a cropped word as input and encodes this image as a sequence of convolutional features. The attention model in between the encoder and the decoder drives, at every step, the focus of attention of the decoder towards a specific part of the sequence of features. Then, an LSTM-based decoder generates a sequence of alpha-numeric symbols as output, one at every time step, terminating when a special stop symbol is output by the LSTM. Below we describe the details of each of the components of the framework.
Encoder: The encoder uses a convolutional neural network to extract a set of features from the image. Specifically, we make use of the CNN model proposed by Jaderberg et al.  for scene text recognition – however we do not use the fully connected layer as a fixed-length representation as it is common in previuos works. Instead, we take the features produced by the last convolutional layer. In this way we can produce a set of feature vectors, each of them linked to a specific spatial location of the image through its corresponding receptive field. This preserves spatial information about the image and reduces model complexity. Through the attention model, the decoder is able to use this spatial information to selectively focus on the most relevant parts of the image at every step.
Thus, given an input image of a cropped word, the encoder generates a set of feature vectors:
where denotes the feature vector corresponding to part of the image. Each corresponds to a spatial location in the image and contains the activations of all feature maps at that location in the last convolutional layer of the CNN.
Attention model: For the attention model, we adapt the soft attention model of  for image captioning, originallly introduced by  for neural machine translation. In  slightly better results are obtained using the hard version of the model that focuses, at every time step, on a single feature vector. However, we argue that, in the case of text recognition, the soft version is more appropriate since a single character will usually span more than one spatial cell of the image corresponding to each of the feature vectors. The soft version of the model can combine several feature vectors with different weights into the final representation.
As shown in figure 2, the attention model generates, at every time step , a vector that will is the input to the LSTM decoder. This vector can be expressed as a weighted combination of the set of feature vectors extracted from the image:
Thus, the vector encodes the relative importance of each part of the image in order to predict the next character for the underlying word. At every time step , and for each location a positive weight is assigned such that . These weights are obtained as the softmax output of a Multi Layer Perpectron (denoted as ) using the set of feature vectors and the hidden state of the LSTM decoder at the previous time step, . More formally:
This model is smooth and differentiable and thus it can be learned using standard back propagation.
Decoder: Our decoder is a Long Short Term Memory (LSTM) network  which produces one symbol from the given symbol set , at every time step. The output of the LSTM is a vector of character probabilities which represents the probability of emitting each of the characters in the symbol set at time . It depends on the output vector of the soft attention model , the hidden state at previous step and the output of the LSTM at previous step . We follow the notation introduced in  where the network is described by:
where is the matrix of weights learned by the network and and are the input, forget, memory, output and hidden state of the LSTM, respectively. In the above definition, denotes the element-wise multiplication and is an embedding of the output character probabilities that is also learned by the network. and denote the activation functions that are applied after the multiplication by the matrix of weights
Finally, to compute the output character probability , a deep output layer is added that takes as input the character probability at the previous step, the current LSTM hidden state, and the current feature vector. The output character probability is:
where , and are the parameters of the deep output layer that are learned using back-propagation.
We use beam search over LSTM outputs to perform word inference. We first introduce the basic procedure, and then describe how we extend it to incorporate language models.
4.1 The basic inference procedure
Once the model is trained, we use a beam search to approximately maximize the following score function over every possible word: :
where is a special symbol signifying the end of a word, which immediately stops the beam search.
The beam search keeps track at every step of the top most probable sequences of characters. For every active branch of the beam search, given the previous character of the sequence, , the output character probability of the LSTM is used to obtain for all characters in the symbol set .
4.2 Incorporating language models
Text is a strongly contextual. There are some strict constraints imposed by the grammar of the language. For example any word in English cannot carry more than two consecutive occurrences of any alphabet letter. Leveraging such knowledge can positively impact the final recognition output. Although the LSTM implicitly learns some dependences between consecutive characters, we show that adding an explicit language model that takes into account longer dependencies gives a significant boost to recognition accuracy.
In this work we use a standard -gram based language model during inference to leverage the language prior. The character -gram model gives probability of a character conditioned on previous characters, where is a parameter of the model:
where, is the number of occurrences of a particular substring in a training corpus.
Finally, the score function in equation 9 can be modified to take the -gram language model into account as:
At every step we fix the parameter of the language model to the number of previously generated characters in order to take into account the longest possible sequence.
4.3 Lexicon-based inference
Although our method is originally designed for unconstrained text recognition, it can also leverage a lexicon whenever available. The use of a lexicon can be integrated by modifying the beam search so that all active sequences that do not correspond to any valid word are automatically removed from the beam. This can be expressed by modifying equation 9 as:
This can be efficiently implemented by storing the lexicon in a trie structure and automatically removing from the beam search any alternative that do not correspond to any partial branch of the trie.
5 Experimental Results
In this section we report on experiments carried out to validate the proposed model for unconstrained scene text recognition.
5.1 Datasets and experimental protocols
We evaluate the performance of the proposed method using the following standard datasets.
Street View Text (SVT) dataset: this dataset contains 647 cropped word images downloaded from Google Street View. In addition to results on totally unconstrained recognition we also report results using the predefined lexicons defined by Wang et al. in . Results with the 50-word lexicon are referred as SVT-50.
ICDAR’03 text dataset: this dataset dataset contains 251 full images and 860 cropped word images . We used the same protocol as [1, 14, 22] and evaluate cropped word images for which the groundtruth text contains only alphanumeric characters and contains at least three characters. As before we report results on both unconstrained and lexicon-based recognition scenarios. We used same strategy as Wang et al.  to create a 50-word lexicon (ICDAR’03-50), while the setup using the lexicon of all ground truth words is referred as ICDAR’03-full.
Evaluation protocol: We use the standard evaluation protocol adopted in most previous work on text recognition in scene images [10, 14, 22]. The accepted metric is word level accuracy in percentage. SVT and ICDAR’03 are used for evaluation. For lexicon-based recognition, we used the same set of 50 for all images in for SVT and ICDAR’03 dataset, as proposed buy Wang et al. .
Implementation details: The CNN encoder used in this work is the Dictnet model by Jaderberg et al. . Their deep convolutional network consists of four convolutional layers and two fully connected layers. In this work we used features from the last convolutional layer. Thus, the feature map used is of size and therefore, the LSTM takes input in the form of .
For lexicon-based recognition when we do not use the lexicon-based inference explained in section 4.3. Instead, we take the output of unconstrained recognition and find the closest word in the lexicon using the Levenshtein edit distance. For lexicon-based inference in unsconstrained datasets (SVT and ICDAR’03) we use the 90k-words lexicon provided by Jaderberg et al. in . The explicit language model is also learned using this 90k word lexicon.
The parameter (see equation 11) to weight the language model with respect to LSTM character probability is empirically established. In our experiments we found the best results with between 0.25 to 0.3
5.2 Baseline performance analysis
In this section we analyze the impact on performance of all the components of the proposed model. We start with a baseline that consists of a simple one layer LSTM network as decoder, without any attention or explicit language model. As we are interested mainly in the impact of the attention model, we use a simple version in which CNN features from the encoder are fed to the LSTM only at the first time step. At every step the output character is determined based on the output of the previous step and the previous hidden state.
In an effort to evaluate each of our contributions, we trained the baseline system and our model with exactly the same training data. For this purpose we randomly sampled one million training samples from the Synth90k  dataset. For validation we used 300,000 samples randomly taken from the same synth90K dataset.
We present the results for each of the component of the framework as described above in Table 1. The attention model outperforms the baseline by a significant margin (around 7%). Also these results confirm the advantage of using an explicit language model in addition to the implicit conditional character probabilities learned by the LSTM model. Using the language model improves accuracy in another 7%. We also see that further constraining the inference wih a dictionary does not improve the result much, probably because the language model is learned from the same 90K dictionary proposed by Jaderberg et al. in .
In comparison with other related works on unconstrained text recognition, it is noteworthy that with only one million training samples our complete framework can learn a better model than Jaderberg et al.  and obtain results that are close to other state-of-the-art methods that are using the whole 9 million sample training dataset (see table 2).
|Baseline (LSTM-no attention)||61.7|
|Proposed (LSTM + attention model)||68.16|
|Proposed (LSTM + attention model + LM)||75.57|
|Proposed (LSTM + attention model+LM+dict)||76.04|
|Methods||SVT-50||SVT||ICDAR’03-50||ICDAR’03-full||ICDAR’03||ABBY Baseline ABBYY ||35.0||-||24.3||-||-||56.0||55.0||-||-|
|Wang et al. ||57.0||-||76.0||62.0||-||Mishra et al. ||73.2||-||-||-||-||81.8||67.8||-||-|
|Novikova et al. ||72.9||-||82.8||-||-||Wang et al. ||70.0||-||-||-||-||90.0||84.0||-||-|
|Goel et al. ||77.3||-||89.7||-||-||Alsharif and Pineau ||74.3||-||-||-||-||93.1||88.6||-||-|
|Almazan et al. ||89.2||-||-||-||-||Lee et al. ||80.0||-||-||-||-||88.0||76.0||-||-|
|Yao et al. ||75.9||-||88.5||80.3||-||Rodriguez-Serrano et al. ||70.0||-||76.1||57.4||-||-||-||-||-|
|Jaderberg et al. ||86.1||-||96.2||91.5||-||Su and Luet al. ||83.0||-||-||-||-||92.0||82.0||-||-|
|Gordo et al. ||90.7||-||-||-||-||*DICT Jaderberg et al. ||95.4||80.7||97.1||92.7||-||98.7||98.6||93.1||90.8|
|Bissacco et al. ||90.4||78.0||-||-||-||Jaderberget al. ||93.2||71.7||95.5||89.6||-||97.8||97.0||89.6||81.|
|Lee et al. ||96.3||80.7||97.9||97.0||88.7||Proposed (LSTM + attention model)||91.7||75.1||93.4||91.0||89.3|
|Proposed (LSTM + attention model + LM)||95.2||80.4||95.7||94.1||92.6|
|Proposed (LSTM + attention model+LM+dict)||95.4||-||96.2||95.7||-|
5.3 Comparison with state of the art
In this section we will compare our result with other related works on scene text recognition. The results of this comparison are shown in table 2. First, we will discuss results on unconstrained text recognition which is the main focus of our work. Then, we will analyze results for lexicon-based recognition.
Unconstrained text recogntion: apart from our method Jaderberg et al. , Lee et al.  and Bissaccco et al.  are the only methods which are capable of performing totally unconstrained recognition of scene text. Among these methods, our visual attention based model performs significantly better than Bissacco et al.  and Jaderberg et al.  in both SVT and ICDAR’03 datasets. Our model also performs as good as Lee et al.  in SVT dataset and outperforms them by 3% in ICDAR’03 dataset, which is significant given the high recognition rates.
If we further compare our model with that of Lee et al. , that also uses different variants of RNN architectures and an attention model on top of CNN features, we find that they use recursive CNN features. They report that this gives an 8% increase in accuracy over the baseline. This success is due to the recurrent nature of the CNN feature which implicitly model the conditional probability of character sequences.using recursive CNN performs better than the traditional convolutional feature. However, the RNN architecture they use improves only 4% over the baseline. In contrast our method rely on traditional CNN features (which can possibly encodes the presence of individual characters as shown in  from lower convolutional layer preserving local spatial characteristics, which reduces the complexity of the model. In addition, as reported in table 1, our combination of LSTM and soft attention model achieves a much larger margin, 14%, over the baseline. Theses results show that a combination of local convolutional features using the context based attention attention performs better or comparable to the previous state-of-the- art results.
Lexicon-based recognition For SVT-50 we can observe that our method obtain a similar result than the best of the methods  specifically designed to work in a lexicon-based scenario. Comparing with methods for unsconstrained text recognition, only the method of Leeet al.  outperforms our best setting. But as we have already discussed, part of this better performance can be explained by the use of the more complex recursive CNN features.
Concerning ICDAR’03-50 and ICDAR’03-full, our results, although do not beat current state of the art are very competitive and comparable to the best performing methods.
Effect of language modeling: The LSTM with attention model implicitly learns some character sequence model as the output at every step is conditioned by the previous character and also the attention is conditioned by the previous hidden state. However, the baseline analysis in 1 shows that apart from this implicit character sequence model, adding an explicit language model helps to improve the performance. This finding is inline with the works of Bissacco et al. , who also uses a static n-gram language model learned by a huge in-house corpus. In our case we used the 90k dictionary as corpus to learn the language probability. To efficiently calculate the language probability we make use of a trie  data structure. In Bissacco et al., authors used a n-gram model with and additionally they also used word probabilities to re-rank the words, for which they learn additional word level language model. The authors analyzed the impact of the language model and they report a reduction of the word error rate by approximately 39.2%. In contrast, our language model is small and improves the recognition accuracy around 4%, a result which is very consistent across the different datasets – see Table 2.
Effect of Training Data As the famous quote of Googleâs Research Director Peter Norvig ”We do not have better algorithms. We just have more data.”. The effectiveness of data can not be denied in today’s machine learning systems. As the real datasets are too small for learning deep networks, Jaderberg et al.  proposed to learn deep models using only a synthetic dataset. This dataset has 9 million cropped word images. In order to analyze the behavior of our attention models with increasing data, we plot in Figure 3 the recognition accuracy in SVT dataset with respect to the number of training samples. As expected, accuracy increased with the training size.
We also analyzed the effect of language model with an increasing number of training samples. Initially with 0.5 million training data the improvement when adding the language model is around 8%. However as we increase the training data size this reliance on language model reduces to around 4%. One reason for this can be that with more data the implicit language model learned by the LSTM is more powerful. The use of an LSTM capable of modeling character sequences can also be an additional reason that explains why the improvement obtained by Bissacco et al. using the language model is greater.
Qualitative Results As the features used to encode the image correspond to different parts of the image, we can visualize the attention at every time step of the network. In Figure 4 we can see those visualizations. We can observe that every time step a part of the image is attended, in a way that roughly mimics the natural reading order. We can also notice that the attention model is performing some kind of implicit character segmentation.
In this paper we proposed an LSTM-based visual attention model for scene text recognition. The model uses convolutional features from a standard CNN as input to an LSTM network that selectively attends to parts of the image at each time step in order to recognize words without resorting to a fixed lexicon. We also propose a modified beam search strategy that is able to incorporate weak language models (-grams) to improve recognition accuracy. Experimental results demonstrate that our approach outperforms or performs comparably to state-of-the-art approaches that use lexicons to constrain inferred output words. Experimental results shows that context plays a important part in case of real data, thus using a explicit language model always helps to improve the result.
In future we can extend the attention model for the text detection task, which will lead to an end-to-end framework for text recognition from images. Moreover, in our current framework convolutional features are taken from one single layer, which can lead to poorer results when the text is either too big or too small. This can be dealt with combining features from multiple layers.
-  J. Almazán, A. Gordo, A. Fornés, and E. Valveny. Word spotting and recognition with embedded attributes. IEEE transactions on pattern analysis and machine intelligence, 36(12):2552–2566, 2014.
-  O. Alsharif and J. Pineau. End-to-end text recognition with hybrid hmm maxout models. CoRR, abs/1310.1811, 2013.
-  D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014.
-  A. Bissacco, M. Cummins, Y. Netzer, and H. Neven. Photoocr: Reading text in uncontrolled conditions. In 2013 IEEE International Conference on Computer Vision, pages 785–792, Dec 2013.
-  C. Case, B. Suresh, A. Coates, and A. Y. Ng. Autonomous sign reading for semantic mapping. In Robotics and Automation (ICRA), 2011 IEEE International Conference on, pages 3297–3303. IEEE, 2011.
-  E. Fredkin. Trie memory. Commun. ACM, 3(9):490–499, Sept. 1960.
-  V. Goel, A. Mishra, K. Alahari, and C. Jawahar. Whole is greater than sum of parts: Recognizing scene text words. In ICDAR, 2013.
-  A. Gordo. Supervised mid-level features for word image representation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015.
-  S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
-  M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Deep structured output learning for unconstrained text recognition. ICLR 2015, 2014.
-  M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Reading text in the wild with convolutional neural networks. CoRR, abs/1412.1842, 2014.
-  M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Synthetic data and artificial neural networks for natural scene text recognition. In NIPS Deep Learning Workshop 2014, 2014.
-  M. Jaderberg, A. Vedaldi, and A. Zisserman. Deep features for text spotting. In ECCV 2014. https://bitbucket.org/jaderberg/eccv2014_textspotting, 2014.
-  C. Lee and S. Osindero. Recursive recurrent nets with attention modeling for OCR in the wild. CoRR, abs/1603.03101, 2016.
-  S. Lee, M. S. Cho, K. Jung, and J. H. Kim. Scene text extraction with edge constraint and text collinearity. In Proc. ICPR, 2010.
-  A. Mishra, K. Alahari, and C. Jawahar. Top-down and bottom-up cues for scene text recognition. In Proc. CVPR, 2012.
-  L. Neumann and J. Matas. Real-time scene text localization and recognition. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 3538–3545. IEEE, 2012.
-  T. Novikova, O. Barinova, P. Kohli, and V. Lempitsky. Large-lexicon attribute-consistent text recognition in natural images. In Proc. ECCV, 2012.
-  J. A. Rodriguez-Serrano, A. Gordo, and F. Perronnin. Label embedding: A frugal baseline for text recognition. International Journal of Computer Vision, 113(3):193–207, 2015.
-  D. L. Smith, J. Field, and E. Learned-Miller. Enforcing similarity constraints with integer programming for better scene text recognition. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 73–80. IEEE, 2011.
-  L. P. Sosa, S. M. Lucas, A. Panaretos, L. Sosa, A. Tang, S. Wong, and R. Young. Icdar 2003 robust reading competitions. In In Proceedings of the Seventh International Conference on Document Analysis and Recognition, pages 682–687. IEEE Press, 2003.
-  K. Wang, B. Babenko, and S. Belongie. End-to-end scene text recognition. In 2011 International Conference on Computer Vision, pages 1457–1464. IEEE, 2011.
-  T. Wang, D. J. Wu, A. Coates, and A. Y. Ng. End-to-end text recognition with convolutional neural networks. In Proc. ICPR, 2012.
-  K. Xu, J. Ba, R. Kiros, K. Cho, A. C. Courville, R. Salakhutdinov, R. S. Zemel, and Y. Bengio. Show, attend and tell: Neural image caption generation with visual attention. CoRR, abs/1502.03044, 2015.
-  C. Yao, X. Bai, B. Shi, and W. Liu. Strokelets: A learned multi-scale representation for scene text recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014.