Abstract: Until recently, most computer vision models remained illiterate, largely ignoring the semantically rich and explicit information contained in scene text. Recent progress in scene text detection and recognition has recently allowed exploring its role in a diverse set of open computer vision problems, e.g. image classification, image-text retrieval, image captioning, and visual question answering to name a few. The explicit semantics of scene text closely requires specific modeling similar to language. However, scene text is a particular signal that has to be interpreted according to a comprehensive perspective that encapsulates all the visual cues in an image. Incorporating this information is a straightforward task for humans, but if we are unfamiliar with a language or scripture, achieving a complete world understanding is impossible (e.a. visiting a foreign country with a different alphabet). Despite the importance of scene text, modeling it requires considering the several ways in which scene text interacts with an image, processing and fusing an additional modality. In this thesis, we mainly focus
on two tasks, scene text-based fine-grained image classification, and cross-modal retrieval. In both studied tasks we identify existing limitations in current approaches and propose plausible solutions. Concretely, in each chapter: i) We define a compact way to embed scene text that generalizes to unseen words at training time while performing in real-time. ii) We incorporate the previously learned scene text embedding to create an image-level descriptor that overcomes optical character recognition (OCR) errors which is well-suited to the fine-grained image classification task. iii) We design a region-level reasoning network that learns the interaction through semantics among salient visual regions and scene text instances. iv) We employ scene text information in image-text matching and introduce the Scene Text Aware Cross-Modal retrieval StacMR task. We gather a dataset that incorporates scene text and design a model suited for the newly studied modality. v) We identify the drawbacks of current retrieval metrics in cross-modal retrieval. An image captioning metric is proposed as a way of better evaluating semantics in retrieved results. Ample experimentation shows that incorporating such semantics into a model yields better semantic results while
requiring significantly less data to converge.

Abstract: Visual Question Answering (VQA) is the task where given an image and a natural language question, the objective is to generate a natural language answer. At the intersection between computer vision and natural language processing, this task can be seen as a measure of image understanding capabilities, as it requires to reason about objects, actions, colors, positions, the relations between the different elements as well as commonsense reasoning, world knowledge, arithmetic skills and natural language understanding. However, even though the text present in the images conveys important semantically rich information that is explicit and not available in any other form, most VQA methods remained illiterate, largely
ignoring the text despite its potential significance. In this thesis, we set out on a journey to bring reading capabilities to computer vision models applied to the VQA task, creating new datasets and methods that can read, reason and integrate the text with other visual cues in natural scene images and documents.
In Chapter 3, we address the combination of scene text with visual information to fully understand all the nuances of natural scene images. To achieve this objective, we define a new sub-task of VQA that requires reading the text in the image, and highlight the limitations of the current methods. In addition, we propose a new architecture that integrates both modalities and jointly reasons about textual and visual features. In Chapter 5, we shift the domain of VQA with reading capabilities and apply it on scanned industry document images, providing a high-level end-purpose perspective to Document Understanding, which has been
primarily focused on digitizing the document’s contents and extracting key values without considering the ultimate purpose of the extracted information. For this, we create a dataset which requires methods to reason about the unique and challenging elements of documents, such as text, images, tables, graphs and complex layouts, to provide accurate answers in natural language. However, we observed that explicit visual features provide a slight contribution in the overall performance, since the main information is usually conveyed within the text and its position. In consequence, in Chapter 6, we propose VQA on infographic images, seeking for document images with more visually rich elements that require to fully exploit visual information in order to answer the questions. We show the performance gap of
different methods when used over industry scanned and infographic images, and propose a new method that integrates the visual features in early stages, which allows the transformer architecture to exploit the visual features during the self-attention operation. Instead, in Chapter 7, we apply VQA on a big collection of single-page documents, where the methods must find which documents are relevant to answer the question, and provide the answer itself. Finally, in Chapter 8, mimicking real-world application problems where systems must process documents with multiple pages, we address the multipage document visual question answering task. We demonstrate the limitations of existing methods, including models specifically designed to process long sequences. To overcome these limitations, we propose
a hierarchical architecture that can process long documents, answer questions, and provide the index of the page where the information to answer the question is located as an explainability measure.

Abstract: The transcription of sheet music into some machine-readable format can be carried out manually. However, the complexity of music notation inevitably leads to burdensome software for music score editing, which makes the whole process
very time-consuming and prone to errors. Consequently, automatic transcription
systems for musical documents represent interesting tools.
Document analysis is the subject that deals with the extraction and processing
of documents through image and pattern recognition. It is a branch of computer
vision. Taking music scores as source, the field devoted to address this task is
known as Optical Music Recognition (OMR). Typically, an OMR system takes an
image of a music score and automatically extracts its content into some symbolic
structure such as MEI or MusicXML.
In this dissertation, we have investigated different methods for recognizing a
single staff section (e.g. scores for violin, flute, etc.), much in the same way as most text recognition research focuses on recognizing words appearing in a given line image. These methods are based in two different methodologies. On the one hand, we present two methods based on Recurrent Neural Networks, in particular, the
Long Short-Term Memory Neural Network. On the other hand, a method based on Sequence to Sequence models is detailed.
Music context is needed to improve the OMR results, just like language models
and dictionaries help in handwriting recognition. For example, syntactical rules
and grammars could be easily defined to cope with the ambiguities in the rhythm.
In music theory, for example, the time signature defines the amount of beats per
bar unit. Thus, in the second part of this dissertation, different methodologies
have been investigated to improve the OMR recognition. We have explored three
different methods: (a) a graphic tree-structure representation, Dendrograms, that
joins, at each level, its primitives following a set of rules, (b) the incorporation of Language Models to model the probability of a sequence of tokens, and (c) graph neural networks to analyze the music scores to avoid meaningless relationships between music primitives.
Finally, to train all these methodologies, and given the method-specificity of
the datasets in the literature, we have created four different music datasets. Two of them are synthetic with a modern or old handwritten appearance, whereas the
other two are real handwritten scores, being one of them modern and the other
old.

Abstract: Searching for text objects in real scene images is an open problem and a very active computer vision research area. A large number of methods have been proposed tackling the text search as extension of the ones from the document analysis field or inspired by general purpose object detection methods. However the general problem of object search in real scene images remains an extremely challenging problem due to the huge variability in object appearance. This thesis builds on top of the most recent findings in the visual attention literature presenting a novel computational model of eye guidance aiming to better describe text object search in real scene images.
First are presented the relevant state-of-the-art results from the visual attention literature regarding eye movements and visual search. Relevant models of attention are discussed and integrated with recent observations on the role of top-down constraints and the emerging need for a layered model of attention in which saliency is not the only factor guiding attention. Visual attention is then explained by the interaction of several modulating factors, such as objects, value, plans and saliency. Then we introduce our probabilistic formulation of attention deployment in real scene. The model is based on the rationale that oculomotor control depends on two interacting but distinct processes: an attentional process that assigns value to the sources of information and motor process that flexibly links information with action.
In such framework, the choice of where to look next is task-dependent and oriented to classes of objects embedded within pictures of complex scenes. The dependence on task is taken into account by exploiting the value and the reward of gazing at certain image patches or proto-objects that provide a sparse representation of the scene objects.
In the experimental section the model is tested in laboratory condition, comparing model simulations with data from eye tracking experiments. The comparison is qualitative in terms of observable scan paths and quantitative in terms of statistical similarity of gaze shift amplitude. Experiments are performed using eye tracking data from both a publicly available dataset of face and text and from newly performed eye-tracking experiments on a dataset of street view pictures containing text. The last part of this thesis is dedicated to study the extent to which the proposed model can account for human eye movements in a low constrained setting. We used a mobile eye tracking device and an ad-hoc developed methodology to compare model simulated eye data with the human eye data from mobile eye tracking recordings. Such setting allow to test the model in an incomplete visual information condition, reproducing a close to real-life search task.

Abstract: There are countless collections of historical documents in archives and libraries that contain plenty of valuable information for historians and researchers. The extraction of this information has become a central task among the Document Analysis researches and practitioners.
There is an increasing interest to digital preserve and provide access to these kind of documents. But only the digitalization is not enough for the researchers. The extraction and/or indexation of information of this documents has had an increased interest among researchers. In many cases, and in particular in historical manuscripts, the full transcription of these documents is extremely dicult due the inherent deciencies: poor physical preservation, dierent writing styles, obsolete languages, etc. Word spotting has become a popular an ecient alternative to full transcription. It inherently involves a high level of degradation in the images. The search of words is holistically
formulated as a visual search of a given query shape in a larger image, instead of recognising the input text and searching the query word with an ascii string comparison. But the performance of classical word spotting approaches depend on the degradation level of the images being unacceptable in many cases . In this thesis we have proposed a novel paradigm called contextual word spotting method that uses the contextual/semantic information to achieve acceptable results whereas classical word spotting does not reach. The contextual word spotting framework proposed in this thesis is a segmentation-based word spotting approach, so an ecient word segmentation is needed. Historical handwritten
documents present some common diculties that can increase the diculties the extraction of the words. We have proposed a line segmentation approach that formulates the problem as nding the central part path in the area between two consecutive lines. This is solved as a graph traversal problem. A path nding algorithm is used to nd the optimal path in a graph, previously computed, between the text lines. Once the text lines are extracted, words are localized inside the text lines using a word segmentation technique from the state of the
art. Classical word spotting approaches can be improved using the contextual information of the documents. We have introduced a new framework, oriented to handwritten documents that present a highly structure, to extract information making use of context. The framework is an ecient tool for semi-automatic transcription that uses the contextual information to achieve better results than classical word spotting approaches. The contextual information is
automatically discovered by recognizing repetitive structures and categorizing all the words according to semantic classes. The most frequent words in each semantic cluster are extracted and the same text is used to transcribe all them. The experimental results achieved in this thesis outperform classical word spotting approaches demonstrating the suitability of the proposed ensemble architecture for spotting words in historical handwritten documents using contextual information.

Abstract: In this work, we develop a generic framework that is able to handle the document retrieval problem in various scenarios such as searching for full page matches or retrieving the counterparts for specific document areas, focusing on their structural similarity or letting their visual resemblance to play a dominant role. Based on the spatial indexing technique, we propose to search for matches of local key-region pairs carrying both structural and visual information from the collection while a scheme allowing to adjust the relative contribution of structural and visual similarity is presented.
Based on the fact that the structure of documents is tightly linked with the distance among their elements, we firstly introduce an efficient detector named Distance Transform based Maximally Stable Extremal Regions (DTMSER). We illustrate that this detector is able to efficiently extract the structure of a document image as a dendrogram (hierarchical tree) of multi-scale key-regions that roughly correspond to letters, words and paragraphs. We demonstrate that, without benefiting from the structure information, the key-regions extracted by the DTMSER algorithm achieve better results comparing with state-of-the-art methods while much less amount of key-regions are employed.
We subsequently propose a pair-wise Bag of Words (BoW) framework to efficiently embed the explicit structure extracted by the DTMSER algorithm. We represent each document as a list of key-region pairs that correspond to the edges in the dendrogram where inclusion relationship is encoded. By employing those structural key-region pairs as the pooling elements for generating the histogram of features, the proposed method is able to encode the explicit inclusion relations into a BoW representation. The experimental results illustrate that the pair-wise BoW, powered by the embedded structural information, achieves remarkable improvement over the conventional BoW and spatial pyramidal BoW methods.
To handle various retrieval scenarios in one framework, we propose to directly query a series of key-region pairs, carrying both structure and visual information, from the collection. We introduce the spatial indexing techniques to the document retrieval community to speed up the structural relationship computation for key-region pairs. We firstly test the proposed framework in a full page retrieval scenario where structurally similar matches are expected. In this case, the pair-wise querying method achieves notable improvement over the BoW and spatial pyramidal BoW frameworks. Furthermore, we illustrate that the proposed method is also able to handle focused retrieval situations where the queries are defined as a specific interesting partial areas of the images. We examine our method on two types of focused queries: structure-focused and exact queries. The experimental results show that, the proposed generic framework obtains nearly perfect precision on both types of focused queries while it is the first framework able to tackle structure-focused queries, setting a new state of the art in the field.
Besides, we introduce a line verification method to check the spatial consistency among the matched key-region pairs. We propose a computationally efficient version of line verification through a two step implementation. We first compute tentative localizations of the query and subsequently employ them to divide the matched key-region pairs into several groups, then line verification is performed within each group while more precise bounding boxes are computed. We demonstrate that, comparing with the standard approach (based on RANSAC), the line verification proposed generally achieves much higher recall with slight loss on precision on specific queries.