Diffusion models have demonstrated exceptional capabilities in generating a broad spectrum of visual content, yet their proficiency in rendering text is still limited: they often generate inaccurate characters or words that fail to blend well with the underlying image. To address these shortcomings, we introduce a novel framework named ARTIST, which incorporates a dedicated textual diffusion model to focus on the learning of text structures specifically. Initially, we pretrain this textual model to capture the intricacies of text representation. Subsequently, we finetune a visual diffusion model, enabling it to assimilate textual structure information from the pretrained textual model. This disentangled architecture design and training strategy significantly enhance the text rendering ability of the diffusion models for text-rich image generation. Additionally, we leverage the capabilities of pretrained large language models to interpret user intentions better, contributing to improved generation quality. Empirical results on the MARIO-Eval benchmark underscore the effectiveness of the proposed method, showing an improvement of up to 15% in various metries.

Document capture applications on smartphones have emerged as popular tools for digitizing documents. For many individuals, capturing documents with their smartphones is more convenient than using dedicated photocopiers or scanners, even if the quality of digitization is lower. However, using a smartphone for digitization can become excessively time-consuming and tedious when a user needs to digitize a document with multiple pages.

The increasing prevalence of Deep Neural Networks (DNNs) in cloud-based services has led to their widespread use through various APIs. However, recent studies reveal the susceptibility of these public APIs to model extraction attacks, where adversaries attempt to create a local duplicate of the private model using data and API-generated predictions. Existing defense methods often involve perturbing prediction distributions to hinder an attacker's training goals, inadvertently affecting API utility. In this study, we extend the concept of digital watermarking to protect DNNs' APIs. We suggest embedding a watermark into the safeguarded APIs; thus, any model attempting to copy will inherently carry the watermark, allowing the defender to verify any suspicious models. We propose a simple yet effective framework to increase watermark transferability. By requiring the model to memorize the preset watermarks in the final decision layers, we significantly enhance the transferability of watermarks. Comprehensive experiments show that our proposed framework not only successfully watermarks APIs but also maintains their utility.

Deep learning based models continue to push the frontier of OCR and Handwritten Text Recognition (HTR) both in terms of increasing accuracy and minimizing the burden of data annotation. For example, the CTC loss [3] allows for line-level training without the explicit spatial positions of the characters and words. Other recent works have further reduced the annotation effort by training systems on entire paragraphs or pages of text without line-level annotations. While these techniques have greatly increased accuracy and significantly reduced the data collection burden, useful outputs beyond just the Unicode character prediction have been lost. Because these systems have fewer explicit segmentation steps, they do not produce character-level or word-level bounding boxes. As a result, downstream applications such as highlighting, redaction tools, and basic document correction, are not provided and must be obtained with other methods.

Sharing annotations encourages feedback, discussion, and knowledge passing among readers and can be beneficial for personal and public use. Prior augmented reality (AR) systems have expanded these benefits to both digital and printed documents. However, despite smartphone AR now being widely available, there is a lack of research about how to use AR effectively for interactive document annotation. We propose Dually Noted, a smartphone-based AR annotation system that recognizes the layout of structural elements in a printed document for real-time authoring and viewing of annotations. We conducted experience prototyping with eight users to elicit potential benefits and challenges within smartphone AR, and this informed the resulting Dually Noted system and annotation interactions with the document elements. AR annotation is often unwieldy, but during a 12-user empirical study our novel structural understanding component allows Dually Noted to improve precise highlighting and annotation interaction accuracy by 13%, increase interaction speed by 42%, and significantly lower cognitive load over a baseline method without document layout understanding. Qualitatively, participants commented that Dually Noted was a swift and portable annotation experience. Overall, our research provides new methods and insights for how to improve AR annotations for physical documents.

Watermarking is a commonly used strategy to protect creators’ rights to digital images, videos and audio. Recently, watermarking methods have been extended to deep learning models – in principle, the watermark should be preserved when an adversary tries to copy the model. However, in practice, watermarks can often be removed by an intelligent adversary. Several papers have proposed watermarking methods that claim to be empirically resistant to different types of removal attacks, but these new techniques often fail in the face of new or better-tuned adversaries. In this paper, we propose the first certifiable watermarking method. Using the randomized smoothing technique, we show that our watermark is guaranteed to be unremovable unless the model parameters are changed by more than a certain $\ell_2$ threshold. In addition to being certifiable, our watermark is also empirically more robust compared to previous watermarking methods.

Decomposing images of document pages into high-level semantic regions (e.g., figures, tables, paragraphs), document object detection (DOD) is fundamental for downstream tasks like intelligent document editing and understanding. DOD remains a challenging problem as document objects vary significantly in layout, size, aspect ratio, texture, etc. An additional challenge arises in practice because large labeled training datasets are only available for domains that differ from the target domain. We investigate cross-domain DOD, where the goal is to learn a detector for the target domain using labeled data from the source domain and only unlabeled data from the target domain. Documents from the two domains may vary significantly in layout, language, and genre. We establish a benchmark suite consisting of different types of PDF document datasets that can be utilized for cross-domain DOD model training and evaluation. For each dataset, we provide the page images, bounding box annotations, PDF files, and the rendering layers extracted from the PDF files. Moreover, we propose a novel cross-domain DOD model which builds upon the standard detection model and addresses domain shifts by incorporating three novel alignment modules: Feature Pyramid Alignment (FPA) module, Region Alignment (RA) module and Rendering Layer alignment (RLA) module. Extensive experiments on the benchmark suite substantiate the efficacy of the three proposed modules and the proposed method significantly outperforms the baseline methods. The project page is at \url{https://github.com/kailigo/cddod}.

We address the problem of open-set recognition, where the goal is to determine if a given sample belongs to one of the classes used for training a model (known classes). The main challenge in open-set recognition is to disentangle open-set samples that produce high class activations from known-set samples. We propose two techniques to force class activations of open-set samples to be low. First, we train a generative model for all known classes and then augment the input with the representation obtained from the generative model to learn a classifier. This network learns to associate high classification probabilities both when image content is from the correct class as well as when the input and the reconstructed image are consistent with each other. Second, we use self-supervision to force the network to learn more informative featues when assigning class scores to improve separation of classes from each other and from open-set samples. We evaluate the performance of the proposed method with recent open-set recognition works across three datasets, where we obtain state-of-the-art results.

State-of-the-art methods for document image classification rely on visual features extracted by deep convolutional neural networks (CNNs). These methods do not utilize rich semantic information present in the text of the document, which can be extracted using Optical Character Recognition (OCR). We first study the performance of state-of-the-art text classification approaches when applied to noisy text obtained from OCR. We then show that fusing this textual information with visual CNN methods produces state-of-the-art results on the RVL-CDIP classification dataset.

Training Handwritten Text Recognition (HTR) models typically requires large amounts of labeled data which often are line or page images with corresponding line-level ground truth (GT) transcriptions. Many digital collections have page-level transcriptions for each image, but the transcription is unformatted, i.e., line breaks are not annotated. Can we train lined-based HTR models using such data? In this work, we present a novel alignment technique for segmenting page-level GT text into text lines during HTR model training. This text segmentation problem is formulated as an optimization problem to minimize the cost of aligning predicted lines with the GT text. Using both simulated and HTR model predictions, we show that the alignment method identifies line breaks accurately, even when the predicted lines have high character error rates (CER). We removed the GT line breaks from the ICDAR-2017 READ dataset and trained a HTR model using the proposed alignment method to predict line breaks on-the-fly. This model achieves comparable CER w.r.t. to the same model trained with the GT line breaks. Additionally, we downloaded an online digital collection of 50K English journal pages (not curated for HTR research) whose transcriptions do not contain line breaks, and achieve 11% CER.

The Connectionist Temporal Classification (CTC) loss function [1] enables end-to-end training of a neural network for sequence-to-sequence tasks without the need for prior alignments between the input and output. CTC is traditionally used for training sequential, single-label problems; each element in the sequence has only one class. In this work, we show that CTC is not suitable for multi-label tasks and we present a novel Multi-label Connectionist Temporal Classification (MCTC) loss function for multi-label, sequence-to-sequence classification. Multi-label classes can represent meaningful attributes of a single element; for example, in Optical Music Recognition (OMR), a music note can have separate duration and pitch attributes. Our approach achieves state-of-the-art results on Joint Handwritten Text Recognition and Name Entity Recognition, Asian Character Recognition, and OMR.

Not all languages and domains of handwriting have large labeled datasets available for training handwriting recognition (HWR) models. One way to address this problem is to leverage high resource languages to help train models for low resource languages. In this work, we adapt HWR models trained on a source language to a target language that uses the same writing script. We do so using only labeled data in the source language, unlabeled data in the target language, and a language model in the target language. The language model is used to produce target transcriptions to allow regular example based training. Using this approach we demonstrate improved transferability among French, English, and Spanish languages using both historical and modern handwriting datasets.

We introduce two data augmentation and normalization techniques, which, used with a CNN-LSTM, significantly reduce Word Error Rate (WER) and Character Error Rate (CER) beyond best-reported results on handwriting recognition tasks. (1) We apply a novel profile normalization technique to both word and line images. (2) We augment existing text images using random perturbations on a regular grid. We apply our normalization and augmentation to both training and test images. Our approach achieves low WER and CER over hundreds of authors, multiple languages and a variety of collections written centuries apart. Image augmentation in this manner achieves state-of-the-art recognition accuracy on several popular handwritten word benchmarks.