This paper presents final results of ICDAR 2019 Scene Text Visual Question Answering competition (ST-VQA). ST-VQA introduces an important aspect that is not addressed by any Visual Question Answering system up to date, namely the incorporation of scene text to answer questions asked about an image. The competition introduces a new dataset comprising 23,038 images annotated with 31,791 question / answer pairs where the answer is always grounded on text instances present in the image. The images are taken from 7 different public computer vision datasets, covering a wide range of scenarios. The competition was structured in three tasks of increasing difficulty, that require reading the text in a scene and understanding it in the context of the scene, to correctly answer a given question. A novel evaluation metric is presented, which elegantly assesses both key capabilities expected from an optimal model: text recognition and image understanding. A detailed analysis of results from different participants is showcased, which provides insight into the current capabilities of VQA systems that can read. We firmly believe the dataset proposed in this challenge will be an important milestone to consider towards a path of more robust and general models that can exploit scene text to achieve holistic image understanding.

This paper explores the possibilities of image style transfer applied to text maintaining the original transcriptions. Results on different text domains (scene text, machine printed text and handwritten text) and cross-modal results demonstrate that this is feasible, and open different research lines. Furthermore, two architectures for selective style transfer, which means transferring style to only desired image pixels, are proposed. Finally, scene text selective style transfer is evaluated as a data augmentation technique to expand scene text detection datasets, resulting in a boost of text detectors performance. Our implementation of the described models is publicly available.

Cross-modal retrieval methods have been significantly improved in last years with the use of deep neural networks and large-scale annotated datasets such as ImageNet and Places. However, collecting and annotating such datasets requires a tremendous amount of human effort and, besides, their annotations are limited to discrete sets of popular visual classes that may not be representative of the richer semantics found on large-scale cross-modal retrieval datasets. In this paper, we present a self-supervised cross-modal retrieval framework that leverages as training data the correlations between images and text on the entire set of Wikipedia articles. Our method consists in training a CNN to predict: (1) the semantic context of the article in which an image is more probable to appear as an illustration, and (2) the semantic context of its caption. Our experiments demonstrate that the proposed method is not only capable of learning discriminative visual representations for solving vision tasks like classification, but that the learned representations are better for cross-modal retrieval when compared to supervised pre-training of the network on the ImageNet dataset.

Word-spotting methods based on the Bag-of-Visual-Words framework have demonstrated a good retrieval performance even when used in a completely unsupervised manner. Although unsupervised approaches are suitable for large document collections due to the cost of acquiring labeled data, these methods also present some drawbacks. For instance, having to train a suitable "codebook" for a certain dataset has a high computational cost. Therefore, in this paper we present a database agnostic codebook which is trained from synthetic data. The aim of the proposed approach is to generate a codebook where the only information required is the type of script used in the document. The use of synthetic data also allows to easily incorporate semantic information in the codebook generation. So, the proposed method is able to determine which set of codewords have a semantic representation of the descriptor feature space. Experimental results show that the resulting codebook attains a state-of-the-art performance while having a more compact representation.

In this paper, we explore the use of second-order derivatives to detect text lines on handwritten document images. Taking advantage that the second derivative gives a minimum response when a dark linear element over a bright background has the same orientation as the filter, we use this operator to create a map with the local orientation and strength of putative text lines in the document. Then, we detect line segments by selecting and merging the filter responses that have a similar orientation and scale. Finally, text lines are found by merging the segments that are within the same text region. The proposed segmentation algorithm, is learning-free while showing a performance similar to the state of the art methods in publicly available datasets.

In this paper, we present an incremental frame-work for extracting information fields from administrative documents. First, we demonstrate some limits of the existing state-of-the-art methods such as the delay of the system efficiency. This is a concern in industrial context when we have only few samples of each document class. Based on this analysis, we propose a hybrid system combining incremental learning by means of itf-df statistics and a-priori generic models. We report in the experimental section our results obtained with a dataset of real invoices.

In this paper we present a segmentation-free system for reading text in natural scenes. A CNN architecture is trained in an end-to-end manner, and is able to directly output readings without any explicit text localization step. In order to validate our proposal, we focus on the specific case of reading utility meters. We present our results in a large dataset of images acquired by different users and devices, so text appears in any location, with different sizes, fonts and lengths, and the images present several distortions such as dirt, illumination highlights or blur.

The ICDAR Robust Reading Competition (RRC), initiated in 2003 and re-established in 2011, has become a de-facto evaluation standard for robust reading systems and algorithms. Concurrent with its second incarnation in 2011, a continuous effort started to develop an on-line framework to facilitate the hosting and management of competitions. This paper outlines the Robust Reading Competition Annotation and Evaluation Platform, the backbone of the competitions. The RRC Annotation and Evaluation Platform is a modular framework, fully accessible through on-line interfaces. It comprises a collection of tools and services for managing all processes involved with defining and evaluating a research task, from dataset definition to annotation management, evaluation specification and results analysis. Although the framework has been designed with robust reading research in mind, many of the provided tools are generic by design. All aspects of the RRC Annotation and Evaluation Framework are available for research use.

In this paper we present an automatic method for separating static and variable content from administrative document images. An alignment approach is able to unsupervisedly build probabilistic templates from a set of examples of the same document kind. Such templates define which is the likelihood of every pixel of being either static or variable content. In the extraction step, the same alignment technique is used to match an incoming image with the template and to locate the positions where variable fields appear. We validate our approach on the public NIST Structured Tax Forms Dataset.

In this paper we present the LSDE string representation and its application to handwritten word spotting. LSDE is a novel embedding approach for representing strings that learns a space in which distances between projected points are correlated with the Levenshtein edit distance between the original strings. We show how such a representation produces a more semantically interpretable retrieval from the user's perspective than other state of the art ones such as PHOC and DCToW. We also conduct a preliminary handwritten word spotting experiment on the George Washington dataset.

Reducing the amount of keypoints used to index an image is particularly interesting to control processing time and memory usage in real-time document image matching applications, like augmented documents or smartphone applications. This paper benchmarks two keypoint selection methods on a task consisting of reducing keypoint sets extracted from document images, while preserving detection and segmentation accuracy. We first study the different forms of keypoint filtering, and we introduce the use of the CORE selection method on keypoints extracted from document images. Then, we extend a previously published benchmark by including evaluations of the new method, by adding the SURF-BRISK detection/description scheme, and by reporting processing speeds. Evaluations are conducted on the publicly available dataset of ICDAR2015 SmartDOC challenge 1. Finally, we prove that reducing the original keypoint set is always feasible and can be beneficial not only to processing speed but also to accuracy.

In this paper, we propose a new feature vector, named DElaunay TRIangulation-based Features (DETRIF), for real-time camera-based document image retrieval. DETRIF is computed based on the geometrical constraints from each pair of adjacency triangles in delaunay triangulation which is constructed from centroids of connected components. Besides, we employ a hashing-based indexing system in order to evaluate the performance of DETRIF and to compare it with other systems such as LLAH and SRIF. The experimentation is carried out on two datasets comprising of 400 heterogeneous-content complex linguistic map images (huge size, 9800 X 11768 pixels resolution) and 700 textual document images.

All indications show that paper documents will not cede in favour of their digital counterparts, but will instead be used increasingly in conjunction with digital information. An open challenge is how to seamlessly link the physical with the digital -- how to continue taking advantage of the important affordances of paper, without missing out on digital functionality. This paper presents the authors' experience with developing systems for Human-Document Interaction based on augmented document interfaces and examines new challenges and opportunities arising for the document image analysis field in this area. The system presented combines state of the art camera-based document image analysis techniques with a range of complementary technologies to offer fluid Human-Document Interaction. Both fixed and nomadic setups are discussed that have gone through user testing in real-life environments, and use cases are presented that span the spectrum from business to educational applications.

Smartphones are enabling new ways of capture, hence arises the need for seamless and reliable acquisition and digitization of documents, in order to convert them to editable, searchable and a more human-readable format. Current state-of-the-art works lack databases and baseline benchmarks for digitizing mobile captured documents. We have organized a competition for mobile document capture and OCR in order to address this issue. The competition is structured into two independent challenges: smartphone document capture, and smartphone OCR. This report describes the datasets for both challenges along with their ground truth, details the performance evaluation protocols which we used, and presents the final results of the participating methods. In total, we received 13 submissions: 8 for challenge-1, and 5 for challenge-2.

In this paper we propose an effective method aimed at reducing the amount of local descriptors to be indexed in a document matching framework. In an off-line training stage, the matching between the model document and incoming images is computed retaining the local descriptors from the model that steadily produce good matches. We have evaluated this approach by using the ICDAR2015 SmartDOC dataset containing near 25 000 images from documents to be captured by a mobile device. We have tested the performance of this filtering step by using ORB and SIFT local detectors and descriptors. The results show an important gain both in quality of the final matching as well as in time and space requirements.

This paper presents a novel way to generate ground-truth data for the evaluation of mobile document capture systems, focusing on the first stage of the image processing pipeline involved: document object detection and segmentation in low-quality preview frames. We introduce and describe a simple, robust and fast technique based on color markers which enables a semi-automated annotation of page corners. We also detail a technique for marker removal. Methods and tools presented in the paper were successfully used to annotate, in few hours, 24889 frames in 150 video files for the smartDOC competition at ICDAR 2015.

Spatial verification is typically employed to check the spatial consistency among matched local features and to remove outliers. However, when looking for multiple instances of the query within a target image, RANSAC algorithms which are widely applied in many one-to-one matching applications might fail due to the large proportion of “outliers” - correct matches corresponding to other instances. On the other hand, geometrical verification methods are more robust to outliers but usually suffer from high computational costs. In this paper, we introduce a novel two-step line verification method which is more flexible than existing methods and leads to lower computational complexity especially when multiple instances of a query are sought. We study this approach within an information extraction scenario, where the objective is to locate document structures indicative of certain type of information (e.g. different records on invoices).

In this paper, we present a new querying paradigm for handwritten keyword spotting. We propose to represent handwritten word images both by visual and audio representations, enabling a query-by-speech keyword spotting system. The two representations are merged together and projected to a common sub-space in the training phase. This transform allows to, given a spoken query, retrieve word instances that were only represented by the visual modality. In addition, the same method can be used backwards at no additional cost to produce a handwritten text-to-speech system. We present our first results on this new querying mechanism using synthetic voices over the George Washington dataset.

In this paper we conduct a comparative study of local key-point detectors and local descriptors for the specific task of mobile document classification. A classification architecture based on direct matching of local descriptors is used as baseline for the comparative study. A set of four different key-point detectors and four different local descriptors are tested in all the possible combinations. The experiments are conducted in a database consisting of 30 model documents acquired on 6 different backgrounds, totaling more than 36.000 test images.

Mobile document image acquisition is a new trend raising serious issues in business document processing workflows. Such digitization procedure is unreliable, and integrates many distortions which must be detected as soon as possible, on the mobile, to avoid paying data transmission fees, and losing information due to the inability to re-capture later a document with temporary availability. In this context, out-of-focus blur is major issue: users have no direct control over it, and it seriously degrades OCR recognition. In this paper, we concentrate on the estimation of focus quality, to ensure a sufficient legibility of a document image for OCR processing. We propose two contributions to improve OCR accuracy prediction for mobile-captured document images. First, we present 24 focus measures, never tested on document images, which are fast to compute and require no training. Second, we show that a combination of those measures enables state-of-the art performance regarding the correlation with OCR accuracy. The resulting approach is fast, robust, and easy to implement in a mobile device. Experiments are performed on a public dataset, and precise details about image processing are given.

In this paper, we present a word spotting framework that follows the query-by-string paradigm where word images are represented both by textual and visual representations. The textual representation is formulated in terms of character n-grams while the visual one is based on the bag-of-visual-words scheme. These two representations are merged together and projected to a sub-vector space. This transform allows to, given a textual query, retrieve word instances that were only represented by the visual modality. Moreover, this statistical representation can be used together with state-of-the-art indexation structures in order to deal with large-scale scenarios. The proposed method is evaluated using a collection of historical documents outperforming state-of-the-art performances.

In this paper we argue that a key-region detector designed to take into account the special characteristics of document images can result in the detection of less and more meaningful key-regions. We propose a fast key-region detector able to capture aspects of the structural information of the document, and demonstrate its efficiency by comparing against standard detectors in an administrative document retrieval scenario. We show that using the proposed detector results to a smaller number of detected key-regions and higher performance without any drop in speed compared to standard state of the art detectors.

In this paper we present a method for the segmentation of continuous page streams into multipage documents and the simultaneous classification of the resulting documents. We first present an approach to combine the multiple pages of a document into a single feature vector that represents the whole document. Despite its simplicity and low computational cost, the proposed representation yields results comparable to more complex methods in multipage document classification tasks. We then exploit this representation in the context of page stream segmentation. The most plausible segmentation of a page stream into a sequence of multipage documents is obtained by optimizing a statistical model that represents the probability of each segmented multipage document belonging to a particular class. Experimental results are reported on a large sample of real administrative multipage documents.

Recent HMM-based approaches to handwritten word spotting require large amounts of learning samples and mostly rely on a prior segmentation of the document. We propose to use Bag-of-Features HMMs in a patch-based segmentation-free framework that are estimated by a single sample. Bag-of-Features HMMs use statistics of local image feature representatives. Therefore they can be considered as a variant of discrete HMMs allowing to model the observation of a number of features at a point in time. The discrete nature enables us to estimate a query model with only a single example of the query provided by the user. This makes our method very flexible with respect to the availability of training data. Furthermore, we are able to outperform state-of-the-art results on the George Washington dataset.

In this paper we present an incremental framework aimed at extracting field information from administrative document images in the context of a Digital Mail-room scenario. Given a single training sample in which the user has marked which fields have to be extracted from a particular document class, a document model representing structural relationships among words is built. This model is incrementally refined as the system processes more and more documents from the same class. A reformulation of the tf-idf statistic scheme allows to adjust the importance weights of the structural relationships among words. We report in the experimental section our results obtained with a large dataset of real invoices.

In this paper, we present a segmentation-free word spotting method that is able to deal with heterogeneous document image collections. We propose a patch-based framework where patches are represented by a bag-of-visual-words model powered by SIFT descriptors. A later refinement of the feature vectors is performed by applying the latent semantic indexing technique. The proposed method performs well on both handwritten and typewritten historical document images. We have also tested our method on documents written in non-Latin scripts.

In this paper we tackle the problem of document image retrieval by combining a similarity measure between documents and the probability that a given document belongs to a certain class. The membership probability to a specific class is computed using Support Vector Machines in conjunction with similarity measure based kernel applied to structural document representations. In the presented experiments, we use different document representations, both visual and structural, and we apply them to a database of historical documents. We show how our method based on similarity kernels outperforms the usual distance-based retrieval.

In this paper we present a method for organizing and indexing logo digital libraries like the ones of the patent and trademark offices. We propose an efficient queried-by-example retrieval system which is able to retrieve logos by similarity from large databases of logo images. Logos are compactly described by a variant of the shape context descriptor. These descriptors are then indexed by a locality-sensitive hashing data structure aiming to perform approximate k-NN search in high dimensional spaces in sub-linear time. The experiments demonstrate the effectiveness and efficiency of this system on realistic datasets as the Tobacco-800 logo database.

In this paper we present a method for document categorization which processes incoming document images such as invoices or receipts. The categorization of these document images is done in terms of the presence of a certain graphical logo detected without segmentation. The graphical logos are described by a set of local features and the categorization of the documents is performed by the use of a bag-of-words model. Spatial coherence rules are added to reinforce the correct category hypothesis, aiming also to spot the logo inside the document image. Experiments which demonstrate the effectiveness of this system on a large set of real data are presented.

In this paper we present a method to spot both text and graphical symbols in a collection of images of wiring diagrams. Word spotting and symbol spotting methods tend to use the most discriminative features to describe the objects to be located. This fact makes that one can not tackle with textual and symbolic information at the same time. We propose a spotting architecture able to index both words and symbols, inspired in off-the-shelf object recognition architectures. Keypoints are extracted from a document image and a local descriptor is computed at each of these points of interest. The spatial organization of these descriptors validate the hypothesis to find an object (text or symbol) in a certain location and under a certain pose.

In this paper we present a method to locate and recognize graphical symbols appearing in real images. A vectorial signature is defined to describe graphical symbols. It is formulated in terms of accumulated length and angular information computed from polygonal approximation of contours. The proposed method aims to locate and recognize graphical symbols in cluttered environments at the same time, without needing a segmentation step. The symbol signature is tolerant to rotation, scale, translation and to distortions such as weak perspective, blurring effect and illumination changes usually present when working with scenes acquired with low resolution cameras in open environments.