Extracting key information from unstructured document images, such as historical documents, receipts, orders and credit notes, plays an important role in office automation including efficient archiving, compliance checking and so on. Conventional approaches [31, 30, 3, 25] maintain a set of templates, each of which consists of key words and their layouts. Although they can usually accurately extract key information from documents, they are not robust against the partial text recognition errors, which usually occurs. To make matters worse, they cannot generalize to documents from unseen templates, which prohibits them from being used in many real application scenarios.
In this paper, we target at key information extraction with a more challenging setting, where training set and test set have different templates. CloudScan 
modeled the key information extraction as Named Entity Recognition problem via concatenating texts as strings, which are classified as predefined categories such as order ID, invoice number and so on (see Figure1 (a)). Although it can generalize to samples of unseen templates, it degrades greatly when lines are not aligned properly due to non-front image captures. Moreover, it makes full use of pre-context and after-context only in the concatenated strings but not neighboring text regions which are not in the same line. We believe that a robust key information extraction approach should be robust against image views, and utilize all context in the spatial neighborhood but not the same horizontal line only.
To this end, we propose an end-to-end Spatial Dual-Modality Graph Reasoning (SDMG-R) approach for key information extraction. We model the unstructured document images as spatial dual-modality graphs with graph nodes as detected text boxes and graph edges as the spatial relations between these nodes (see Figure 1
(b)). Each node is associated with the textual and visual features which are learned via a recurrent neural network (RNN) and convolutional neural network (CNN) automatically. Features of graph nodes are propagated iteratively along graph edges before classifying into pre-defined key information categories. In this way, SDMG-R makes full use of spatial relations between detected text regions, and their dual modality features. It is independent of document templates, and thus naturally has the potential to extract key information from document images of unseen templates.
Most of previous key information extraction approaches are evaluated on private data only due to the lack of public datasets. Recently, a few datasets such as IEHHR , SROIE , which target at key information extraction, have been emerging. However, their training and test set share many templates, and thus they are unsuited to evaluate the generalization ability of key information extraction methods. To this end, we build a new key information extraction benchmark dubbed WildReceipt. It consists of 25 key information categories, totally about 50000 text boxes, which is about 2 times larger than SROIE. The key information categories in WildReceipt are fine-grained. e.g., they contain “Subtotal value”, “Total value” and “Tax value” categories, all of which are money amount, and it is difficult to distinguish with each other without context. Different from previous scanned images as in SROIE, receipt images in WildReceipt are captured in the wild, most of them are from non-front views and with folds. Therefore, it is more challenging and realistic than previous ones.
We extensively evaluate our proposed SDMG-R on SROIE and WildReceipt. It has been shown that the proposed approach outperforms previous methods with impressive margins. We investigate the factors of the effectiveness of the proposed approach, and find that both the spatial relations and the dual modality features benefit the key information extraction.
The contributions of this paper are as follows:
We propose an effective spatial dual-modality graph reasoning network (dubbed SDMG-R) for key information extraction. To the best of our knowledge, our SDMG-R is the first key information extraction approach which jointly reasons key information categories on textual and visual features of text boxes and their 2-dimensional spatial relationships.
We annotate a new benchmark, named WildReceipt, to facilitate the future research of key information extraction, which is fine-grained, and 2 times bigger than its competitors. It targets at evaluating key information extraction from document images of unseen templates captured in the wild, which is not explored in previous datasets.
We validate the effectiveness of the proposed SDMG-R on two benchmarks, i.e., SROIE and WildReceipt. Our proposed approach outperforms state-of-the-art approaches with impressive margins.
Ii Related Work
Key information extraction has been attracting a large number of researchers from computer vision and multimedia filed. However, most of them conducted experiments on private datasets. Intellix was trained on 8000 and tested on 4000 scanned documents with 10 annotated semantic categories. CloudScan  conducted experiments on scanned UBL invoices. Later, CUTIE annotated Spanish receipt documents captured in the wild including taxi receipts, Meals Entertainment (ME) receipts, and hotel receipts, with 9 different key information classes. However, all the aforementioned datasets are unavailable publicly. Recently, SROIE  consisting of 600 scanned receipts from ICDAR 2019 Robust Reading Challenge is released with 4 categories. Namely, store name, store address, date and total price. Its training set and its corresponding ground truth are publicly available. However, the test set ground truth is not released. We annotate its test set for evaluating key information extraction approaches on it in our experiments. Our WildReceipt targets at facilitating the key information extraction research and will be publicly released. It is about 2 times and 5 times bigger than SROIE in terms of the total image number and the key information category number respectively. Different from previous datasets which contain value categories only, WildReceipt contains both key and value categories such as “Total key” category (e.g., “Total:” and “Total” ) and “Total value” category (e.g., “$10.5” and “$20”). We empirically find that key category classification during training can boost the classification performance of value categories. Our WildReceipt contains fine-grained categories such as “Total value”, “Subtotal value”, and “Tax value”, all of which represent money amount. Moreover, it is captured in the wild, which is more challenging and wider applicable. The detailed comparison between the key information extraction datasets is conducted in Table I.
Key information extraction. Conventional approaches [31, 30, 3, 6] utilized template matching or rule based strategies, and thus performed poorly on documents of unseen templates. Later, CloudScan  modeled the key information extraction problem as NER [28, 20, 5, 24], and concatenated the entire invoice texts as one dimensional sequences without utilizing the two-dimensional spatial layout information. Chargrid 
encoded each document page as a two-dimensional grid of characters to conduct semantic segmentation. It utilized two-dimensional spatial layout information with small neighborhood only, and could not make fully use of nonlocal spatial relations between text regions with long distances due to limited effective receptive filed of convolution neural networks. Recently, VRD  learned graph embedding to summarize the two-dimensional context of text segments in the document, which were further combined with text embedding for entity extraction from visually rich documents. Our proposed SDMG-R models documents as fully-connected graphs with text regions as nodes and two-dimensional spatial relations as edges. Different from VRD, SDMG-R learns both visual features and textual features of text regions, which leads to robustness against text recognition errors. Detailed comparison between recent approaches with our SDMG-R is given in Table II.
|Methods||1d context||2d context||Nonlocal context||Visual features|
Graph neural networks.
Recently, integrating graphs with deep neural networks is an emerging topic in deep learning research. A considerable amount of models have arisen for reasoning on graph-structured data at various tasks, such as classification of graphs[8, 7, 26], classification of nodes of graphs [17, 18], and modeling multi-agent interacting physical systems [17, 13, 33]. Graph neural networks have been widely used in many application such as human action recognition [38, 36], social relationship understanding , object parsing , multi-label image recognition , visual question answer , and fashion retrieval . These work create graphs via modeling the relations between image objects or regions. In contrast, we explore the use of graph to express spatial relations between detected text boxes which are encoded by visual-textual features, and apply it to the field of key information extraction. For each detected text box, it can automatically mine the useful layout structure in its neighborhood.
Given one document image of size , together with detected text regions , where with , , , and being the top-left corner coordinate, the height, the width, and the recognized text string of respectively, the key information extraction aims at classifying each detected text region into one of a predefined category set . We model the key information extraction as the graph node classification problem via jointly making full use of dual modality features. Namely, visual features and textual ones. Our proposed spatial dual-modality graph reasoning model consists of the dual-modality fusion module, the graph reasoning module and the classification module. Figure 2 shows its overall architecture.
Iii-a Dual-Modality Fusion Module
Given one image with the text regions
, we learn a feature vectorto represent each text region via the dual-modality fusion module. The dual-modality fusion module is designed to effectively learn and compose the visual features and textual ones. We extract the visual feature for by RoI Pooling  with its rectangle on the output feature maps of the last layer of one CNN feature extractor. In our experiments, we use U-Net  to instantiate the CNN feature extractor. Besides, we extract the textual feature for by designing a char-level Bi-LSTM . Specifically, we first represent each char in as a one-hot vector with dimension being the cardinality of the char dictionary. is then projected into a lower dimensional space and finally sequentially fed into the Bi-LSTM module to obtain the textual representation for the text region . Formally, we have
where is the projection matrix of textual one-hot vectors. We fuse the visual features and textual features via modeling the interactions between all possible visual and textual feature dimensional pairs, which are easily obtained with Kronecker product as follows:
is the Kronecker product operation.
is one learnable linear transformation andis the fused feature. For simplicity, we ignore the bias term in our paper. The number of learnable parameters in Equation (2) grows linearly with the dimension of the visual features, that of the textual ones, and that of the fused representations, which results in heavy memory and computation overheads. To reduce the memory and computation complexity, we first reformulate Equation (2) into tensor form:
where is the block-diagonal core tensor with being the block number and being the block size, , and . Usually, we set , , and to one small constant. Thus, the parameter number decomposed tensor in Equation (4) is greatly smaller than that of original tensor in Equation (3). i.e., as shown in our experiments.
We also implement alternative fusion schemes in our experiments for comparison.
LinearSum. The visual features and textual features are linearly projected into one common space via one three-layer MLP, and then element-wise added as the fused representation of .
ConcatMLP. The visual features and textual features are concatenated, followed by one three-layer MLP.
Iii-B Graph Reasoning Module
We model the document images as graphs , where with being the feature vector of the text node , and with being the edge weight between the node and the node .
We encode the spatial relation between and via one dynamic attention mechanism. We first define the spatial relation between node and as follows:
where , and are the horizontal distance and the vertical one between the two text boxes and respectively. is one normalization constant, and is the concatenation operation. The spatial position relation between two text boxes plays a critical role in key information extraction. encodes the relative spatial position distance between node and . The first term and latter two ones in Equation (8) encode the aspect ratio of and relative shape information respectively.
Inspired by , we embed the spatial information between text boxes into the edge weight as follows:
where is one linear transformation which embeds the spatial relation information into a -dimensional representation. is the normalization operation, which is introduced to stabilize the training procedure. is the concatenated representations of , and the normalized spatial relation embedding. is one MLP which transforms into the scalar .
Graph reasoning. We iteratively refine the features of the proposed spatial dual-modality graph times as follows:
where indicates the feature of the graph node at time step . is the normalized graph edge weight at time step . is a linear transformation at time step . is the concatenated representation of , and the normalized spatial relation embedding at time step as described in Equation (11).
is the ReLU nonlinear activation.is the learnable normalized edge weight between node and at time step . It is given by
From Equation (14), the edge weights of the proposed graph change dynamically during reasoning from one iteration to another.
The final output of the iterative reasoning module is fed to the classification module to classify each text region into one of key information categories. Formally, our loss is defined as
where is the key information category ground truth.
Str nm key
Str nm value
Str addr key
Str addr value
Prod item key
Prod item value
Prod qty key
Prod qty value
Prod price key
Prod price value
Iv-a Data Collection
We selected receipts to benchmark key information extraction as SROIE  due to the following reasons: (1) receipts are anonymous, and suitable for public release without private information leak; (2) receipts are of varied templates since different companies usually have different templates. Thus, it is suitable for evaluating key information extraction from document images of unseen templates; (3) receipts are widely available and easy to collect; (4) extracting key information from receipts have many applications such as bookkeeping, and reimbursement.
We collected and annotated WildReceipt in the following procedure.
Data collection. We searched receipt images on search engines with related key-words, such as receipt, invoice and so on. We downloaded about 4300 document images.
Data cleaning. We removed images which have multiple receipts inside, are not receipts, unreadable, incomplete, or non-English manually.
Data annotation. We first labelled the text bounding boxes and their corresponding texts, and then labelled each bounding box to one of 25 key information categories (see Figure 3). These annotations were done by 6 experts.
The receipt images in WildReceipt we selected are captured in the wild. They are of non-front views and possibly with folds as shown in Figure 3. Therefore, WildReceipt is much more challenging than previous key information extraction benchmarks which focus on scanned documents only.
The WildReceipt dataset consists of 1740 receipt images, 68975 text bounding boxes. Each image has average about 39 text bounding boxes. Table III lists the annotation numbers of all 25 key information categories. In the 25 key information categories, 12 categories are keys and 12 categories are their corresponding values, and 1 category indicates others. As there are many variants for one type of key, e.g., “Address”, “address”, and “Add.” all indicate the key category “Str addr key”. We believe that accurately identifying key categories can benefit greatly key information extraction, which is validated in our experiments. WildReceipt is 2 times and 3 times larger than SROIE  in terms of the image number and the category number. Besides, it contains fine-grained key information categories. e.g., “Product price value”, “Tax value”, “Tips value” and “Total value” all are related with the amount of money, and difficult to distinguish with each other by their own textual or visual features without context information.
Iv-C Evaluation Protocol
We randomly sampled 1268 and 472 images without replacement for training and test respectively. During sampling, we made sure these two sets had different templates according to store names and near-duplicated image retrieval. In this way, the templates in test set are unseen in the training set. Therefore, WildReceipt is suitable for evaluating key information extraction from document images of unseen templates. Table IV lists the statistics for the training set and the test set in WildReceipt.
Performances on WildReceipt are evaluated by score as . The averaged score over value categories is finally reported. WildReceipt will be publicly released to facilitate future research and fair comparison on key information extraction.
In this section, the proposed approach SDMG-R is extensively evaluated on SROIE and WildReceipt. We first introduce the implementation details. Then, SDMG-R is compared with state-of-the-art key information extraction approaches quantitatively. Finally, we investigate the effectiveness of each component of our proposed method by ablation study.
V-a Implementation Details
Our implementation is based on PyTorch. Our models are trained on 1 NVIDIA Titan X GPUs with 12 GB memory. During training, we randomly crop images with probability 0.5 while keeping all text boxes not cutting out. During test, we do not crop images. In both training and test, all images are resized to, and their text boxes are resized proportionally before being fed into the network. The whole network is trained from scratch with default initializer of PyTorch using Adam optimizer 
. We use a batch size of 4 during training. Maximum epoch number is set to. The learning rate is set to initially. It is decreased via after 40 and 50 epochs.
The cardinality of our dictionary is 91 (i.e., ). It contains 0-9 digital, a-z and A-Z letters, and special characters which are greatly related to key information categories. They are “/”, “”, “.”, “$”, “€”, “₤”, “¥”, “:”, “-”, “*”, “#”, “(”, “)”, “%”, “@”, “!”, “”’, “&”, “=”, “¿”, “+”, “””, “”, “?”, “¡”, “[”, “]”, and “_”. All other characters in texts are set to one token “unkown”. The one-hot char encoding vectors are projected to a 32-dimensional space (i.e., ). The dimension of the hidden vector of Bi-LSTM is set to . As for visual modality, we adopt the U-Net  as our visual feature extractor, and extract visual features on its last convolutional output feature maps, followed by a dimensional reduction to . Thus, we have . In the block tensor decomposition module, we set and . We set the graph node feature representation dimension to (i.e., ). The normalization constant is set to (i.e.,) in Equation (7). The 5-dimensional edge features are embedded into one 256-dimensional space (i.e., ). The MLP ( in Equation (12)) is of one two layers with one ReLU between them. Its hidden dimension is . The graph reasoning iteration number is set to 2 (i.e., .) except otherwise noted.
V-B Comparison with State-of-the-art Methods
We compare our proposed SDMG-R with two state-of-the-art approaches and their variants. Specially, we evaluate the following methods:
Chargrid . It models documents as two-dimensional grids of characters, which are fed into a fully convolutional neural network to predict segmentation masks.
Chargrid-UNet. For fair comparison, we also use U-Net as Chargrid’s backbone while keeping other unchanged. We name Chargrid with this setting as Chargrid-UNet.
VRD . It models documents with text bounding boxes as graphs, which are then fed into one CRF.
|Method||Str nm||Str addr||Tel||Date||Time||Prod item||Prod qty||Prod price||Subtotal||Tax||Tips||Total||Avg|
|Method||Str nm||Str addr||Tel||Date||Time||Prod item||Prod qty||Prod price||Subtotal||Tax||Tips||Total||Avg|
|Method||Str nm||Str addr||Tel||Date||Time||Prod item||Prod qty||Prod price||Subtotal||Tax||Tips||Total||Avg.|
We compare our proposed method with its counterparts in Table V. It has been shown that our SDMG-R outperforms all its competitors with impressive margins. Specifically, SDMG-R achieves 11.8%, 9.7%, and 3.0% absolute improvements in terms of score averaged on 12 value categories on WildReceipt compared with Chargrid, Chargrid-UNet, and VRD respectively. Moreover, SDMG-R achieves best score for 10 out of 12 categories. Our SDMG-R is greatly superior than Chargrid-UNet. We believe it is because of the long range dependence between texts learned via graphs. Compared with VRD, the performance gain of the SDMG-R attributes to our proposed U-Net based visual modality and Kronecker product based modality fusion. For the categories “Time” and “Prod qty”, our proposed SDMG-R and VRD are comparable.
Since in real applications, text boxes and texts are usually obtained by OCR engines, which might introduce text detection and recognition errors. To evaluate how those errors affect the performance of the key information extraction, we employ Google OCR API111https://cloud.google.com/vision/docs/ocr to detect and recognize texts of WildReceipt. For each detected text box, we label its key information category as that of the ground truth text region of maximum IOU with it. We compare our SDMG-R with state-of-the-art methods when texts are recognized using the OCR engine given ground truth text boxes in Table VI. Again, our proposed SDMG-R achieves the best averaged score. Moreover, it obviously outperforms its competitors in 10 out of 12 categories. Comparing Table V and Table VI, we observe that there are about perform drop ( v.s. ) in terms of averaged score if texts are recognized by the OCR engine. It is reasonable as some of texts, especially, characters which are closely related to some specific key information categories such as “$”, “€”, “₤”, “¥” are misrecognized via the OCR engine, which results in noisy signals and poor discriminative representations. To move forward, we compare our method with other methods in the case that both text boxes and texts are predicted by the OCR engine in Table VII. It has been shown that our proposed SDMG-R outperforms Chargird, Chargird-UNet and VRD with impressive margins. Note that there exists mismatching between detected text boxes and ground truth boxes. e.g., one detected text boxes might overlap with multiple ground truth ones or one ground truth text box might overlap with multiple detected ones. Directly matching text boxes with ground truth ones with maximum IOU might introduce noisy signals, which results in further perform drop. However, our method is still superior than its counterparts, which validates its robustness against noises.
We also compare our method with other start-of-the-art approaches on the dataset SROIE in Table VIII. Similar to WildReceipt, our SDMG-R obviously performs better than others. Specially, it absolutely improves the scores of Chargrid, Chargrid-UNet, and VRD by , and respectively. It has demonstrated the superiority of our SDMG-R on scanned document images.
|w/o textual features||80.1|
|w/o visual features||86.4|
|w/o spatial relation||81.8|
|w/o graph reasoning||77.2|
|w/o key category classification||84.3|
V-C Ablation Studies
We perform detailed ablation studies on WildReceipt to investigate the effectiveness of each component of our proposed SDMG-R.
Effects of visual and textual features. In Table IX, SDMG-R decreases absolutely by on WildReceipt in terms of score when without textual features. Similarly, it decreases absolutely by when without visual features. It has been shown that both textual and visual features, especially, textual features, contribute the key information extraction greatly.
Effects of spatial relation. To cancel out the spatial relation, we set the edge weights in Equation (10) for all graph node pairs . SDMG-R decreases its score to on WildReceipt. We have observed that spatial relations between two text boxes play an important role in key information extraction and can boost its performance obviously.
Effects of graph reasoning. Without graph reasoning, we directly conduct classification over the fused visual and textual features, resulting in great performance degradation. Namely, absolute score drop on WildReceipt. It suggests that message propagation between text regions can refine their representations so that they can be correctly classified into their corresponding key information categories.
Effects of key category classification. In our WildReceipt dataset, we also annotate key categories such as “Str nm key”, although only the information of value categories needs to be extracted in real application scenarios. However, we experimentally find that key category classification can help the value category classification. As shown in Table IX, without it, our SDMG-R decreases absolutely by 4.4%.
Effects of graph reasoning iteration number. Our SDMG-R obtains the averaged score of , and when the graph reasoning iteration number is set to , and respectively. It achieves the best performance when , and is overfitted when . We set in our experiments.
Effects of dual modality fusion module. Dual modality fusion module is the core component to fuse visual and textual features. We compare our module with its counterparts LinearSum and ConcatMLP as described in Section III-A. For fair comparison, we enumerate the hidden dimension () of the MLP in LinearSum and ConcatMLP, and the block size (, , ) and the block num () in our proposed dual modality fusion module, and report their corresponding results in Table X. We can observe that ConcatMLP and LinearSum achieve their best results with (or ), and respectively while our method with and . It has been shown that our proposed dual modality fusion module is very effective, and obviously outperforms its alternative methods. Namely, LinerSum and ConcatMLP.
To better understand how our SDMG-R learns the spatial relations between the text regions, we visualize the learned edge weights in Figure 4. Interestingly, it can highlight the edges between two semantically-related text regions even they are with long spatial distances. e.g., the edges between “Total value”, “Subtotal value” and “Prod item value” (top left), those between “Total value”, “Total key”, and “Tax key” (top right), those between “Subtotal value” and “Subtotal key” (bottom left), and those between “Subtotal key”, “Tax key”, “Total key” and “Subtotal value” (bottom right). Compared with the first GCL (the left column), the second GCL can learn more helpful spatial relations to identity the key information categories of text regions. e.g., “TOTAL” on the left of “$30473.00” highly indicates “$30473.00” is one instance of “Total value” in the top right subfigure. “TOTAL:” and “TAXES:” under the “SUBTOTAL:” indicate “SUBTOTAL:” is one instance of “Subtotal key” in the bottom right subfigure.
In this paper, we have proposed a novel spatial dual-modality graph reasoning model (termed SDMG-R) for key information extraction from unstructured documents. We have introduced Kronecker product approximated via the block diagonal tensor decomposition to fuse the visual and textual features. SDMG-R naturally learns spatial relations between text regions via dynamical attentions in its graph reasoning module. We have validated the effectiveness of each component of the proposed SDMG-G by extensive experiments. Moreover, a new large key information extraction dataset, named WildReceipt, has been annotated to evaluate the model performance of the key information extraction on document of unseen templates. It is fine grained and captured in the wild, and thus more challenging and realistic than previous public datasets. It will be publicly released for facilitating future research. Experimental results on both SROIE and our WildReceipt databases have shown that our proposed SDMG-R consistently outperforms start-of-the-art key information extraction methods with impressive margins.
-  (2019) BLOCK: Bilinear Superdiagonal Fusion for Visual Question Answering and Visual Relationship Detection. In AAAI, Vol. 33, pp. 8102–8109. Cited by: §II, §III-A.
-  (2017) MUTAN: Multimodal Tucker Fusion for VQA. In ICCV, pp. 2612–2620. Cited by: §II.
-  (2003) Analysis and understanding of multi-class invoices. Document Analysis and Recognition 6 (2), pp. 102–114. Cited by: §I, §II.
Recurrent Attentional Reinforcement Learning for Multi-label Image Recognition. In AAAI, Cited by: §II.
-  (2016) Named entity recognition with bidirectional lstm-cnns. Transactions of the Association for Computational Linguistics 4, pp. 357–370. Cited by: §II.
-  (2018) Field extraction by hybrid incremental and a-priori structural templates. In 2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pp. 251–256. Cited by: §II.
-  (2016) Discriminative embeddings of latent variable models for structured data. In ICML, pp. 2702–2711. Cited by: §II.
-  (2015) Convolutional networks on graphs for learning molecular fingerprints. In NeurIPS, pp. 2224–2232. Cited by: §II.
-  (2018) Chargrid: Towards Understanding 2D Documents. In EMNLP, External Links: Cited by: §II, 1st item.
-  (2017) ICDAR2017 competition on information extraction in historical handwritten records. In ICDAR, Vol. 1, pp. 1389–1394. Cited by: §I.
-  (2016) Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding. In EMNLP, pp. 457–468. Cited by: §II.
-  (2015) Fast r-cnn. In ICCV, Cited by: §III-A.
-  (2017) Vain: attentional multi-agent predictive modeling. In NeurIPS, pp. 2701–2711. Cited by: §II.
-  (2019) ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction. Cited by: §I, §II, §IV-A, §IV-B, §IV-C.
-  (2017) Hadamard Product for Low-rank Bilinear Pooling. In ICLR, pp. 1–14. Cited by: §II.
-  (2014) Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980. Cited by: §V-A.
-  (2018) Neural relational inference for interacting systems. arXiv preprint arXiv:1802.04687. Cited by: §II.
-  (2016) Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. Cited by: §II.
-  (2019) Fashion Retrieval via Graph Reasoning Networks on a Similarity Pyramid. In ICCV, Cited by: §II.
-  (2016) Neural architectures for named entity recognition. arXiv preprint arXiv:1603.01360. Cited by: §II.
-  (2016) Semantic Object Parsing with Graph LSTM. In ECCV, External Links: Cited by: §II.
-  (2019) Graph Convolution for Multimodal Information Extraction from Visually Rich Documents. In NAACL, pp. 32–39. Cited by: §II, 3rd item.
-  (2016) Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (Eds.), pp. 4898–4906. Cited by: §II.
-  (2016) End-to-end sequence labeling via bi-directional lstm-cnns-crf. arXiv preprint arXiv:1603.01354. Cited by: §II, §III-A.
-  (2011) A probabilistic approach to printed document understanding. International Journal on Document Analysis and Recognition (IJDAR) 14 (4), pp. 335–347. Cited by: §I.
-  (2016) Progressively parsing interactional objects for fine grained action detection. In CVPR, pp. 1020–1028. Cited by: §II.
-  (2017) CloudScan-a configuration-free invoice analysis system using recurrent neural networks. In ICDAR, Vol. 1, pp. 406–413. Cited by: §I, §II, §II.
-  (2017) Cross-sentence n-ary relation extraction with graph lstms. Transactions of the Association for Computational Linguistics 5, pp. 101–115. Cited by: §II.
-  (2015) U-net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), Vol. 9351, pp. 234–241. Cited by: §III-A, §V-A.
-  (2013) Field extraction from administrative documents by incremental structural templates. In 2013 12th International Conference on Document Analysis and Recognition, pp. 1100–1104. Cited by: §I, §II.
-  (2013) Intellix - End-User Trained Information Extraction for Document Archiving. In ICDAR, Cited by: §I, §II, §II.
-  (2003) Video google: a text retrieval approach to object matching in videos.. In ICCV, pp. 1470–1477. Cited by: §IV-C.
Learning multiagent communication with backpropagation. In NeurIPS, pp. 2244–2252. Cited by: §II.
-  (2017) Graph-Structured Representations for Visual Question Answering. In CVPR, pp. 1–9. Cited by: §II.
-  (2018) Graph attention networks. In International Conference on Learning Representations, Cited by: §III-B.
-  (2018) Videos as space-time region graphs. In ECCV, pp. 399–417. Cited by: §II.
Deep Reasoning with Knowledge Graph for Social Relationship Understanding. In IJCAI, External Links: Cited by: §II.
-  (2018) Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, Cited by: §II.
-  (2018) Learning compact recurrent neural networks with block-term tensor decomposition. In CVPR, pp. 9378–9387. Cited by: §III-A.