Cha Zhang

Junlong Li, Yiheng Xu, Tengchao Lv et al. · microsoft-research

Image Transformer has recently achieved significant progress for natural image understanding, either using supervised (ViT, DeiT, etc.) or self-supervised (BEiT, MAE, etc.) pre-training techniques. In this paper, we propose \textbf{DiT}, a self-supervised pre-trained \textbf{D}ocument \textbf{I}mage \textbf{T}ransformer model using large-scale unlabeled text images for Document AI tasks, which is essential since no supervised counterparts ever exist due to the lack of human-labeled document images. We leverage DiT as the backbone network in a variety of vision-based Document AI tasks, including document image classification, document layout analysis, table detection as well as text detection for OCR. Experiment results have illustrated that the self-supervised pre-trained DiT model achieves new state-of-the-art results on these downstream tasks, e.g. document image classification (91.11 $\rightarrow$ 92.69), document layout analysis (91.0 $\rightarrow$ 94.9), table detection (94.23 $\rightarrow$ 96.55) and text detection for OCR (93.07 $\rightarrow$ 94.29). The code and pre-trained models are publicly available at \url{https://aka.ms/msdit}.

Zineng Tang, Ziyi Yang, Guoxin Wang et al. · microsoft-research

We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, image, and layout modalities together with varied task formats, including document understanding and generation. UDOP leverages the spatial correlation between textual content and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and layout modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high-quality neural document editing and content customization. Our method sets the state-of-the-art on 8 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and websites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark.

Tengchao Lv, Yupan Huang, Jingye Chen et al. · microsoft-research

The automatic reading of text-intensive images represents a significant advancement toward achieving Artificial General Intelligence (AGI). In this paper we present KOSMOS-2.5, a multimodal literate model for machine reading of text-intensive images. Pre-trained on a large-scale corpus of text-intensive images, KOSMOS-2.5 excels in two distinct yet complementary transcription tasks: (1) generating spatially-aware text blocks, where each block of text is assigned spatial coordinates within the image, and (2) producing structured text output that captures both style and structure in markdown format. This unified multimodal literate capability is achieved through a shared decoder-only autoregressive Transformer architecture and task-specific prompts. Building on this foundation, we fine-tune KOSMOS-2.5 for document understanding tasks, resulting in a document understanding generalist named KOSMOS-2.5-CHAT. Additionally, a large corpus of 357.4 million document pages spanning diverse domains was curated for pre-training. We evaluate KOSMOS-2.5 on two newly proposed benchmarks, OCREval and MarkdownEval, for document-level text recognition and image-to-markdown generation, demonstrating impressive literate capabilities comparable to GPT-4o. KOSMOS-2.5-CHAT achieves performance comparable to other state-of-the-art generalists that are five times larger (1.3B vs. 7B) across nine text-rich visual question answering benchmarks. Models and code have been available at \url{https://aka.ms/kosmos25}.

Jingye Chen, Tengchao Lv, Lei Cui et al. · microsoft-research

The surge of pre-training has witnessed the rapid development of document understanding recently. Pre-training and fine-tuning framework has been effectively used to tackle texts in various formats, including plain texts, document texts, and web texts. Despite achieving promising performance, existing pre-trained models usually target one specific document format at one time, making it difficult to combine knowledge from multiple document formats. To address this, we propose XDoc, a unified pre-trained model which deals with different document formats in a single model. For parameter efficiency, we share backbone parameters for different formats such as the word embedding layer and the Transformer layers. Meanwhile, we introduce adaptive layers with lightweight parameters to enhance the distinction across different formats. Experimental results have demonstrated that with only 36.7% parameters, XDoc achieves comparable or even better performance on a variety of downstream tasks compared with the individual pre-trained models, which is cost effective for real-world deployment. The code and pre-trained models will be publicly available at \url{https://aka.ms/xdoc}.

Hai Pham, Guoxin Wang, Yijuan Lu et al. · cmu

Despite several successes in document understanding, the practical task for long document understanding is largely under-explored due to several challenges in computation and how to efficiently absorb long multimodal input. Most current transformer-based approaches only deal with short documents and employ solely textual information for attention due to its prohibitive computation and memory limit. To address those issues in long document understanding, we explore different approaches in handling 1D and new 2D position-aware attention with essentially shortened context. Experimental results show that our proposed models have advantages for this task based on various evaluation metrics. Furthermore, our model makes changes only to the attention and thus can be easily adapted to any transformer-based architecture.

Kilho Son, Paul Hsu, Cha Zhang et al.

Conversational search has become increasingly important in retrieval-augmented generation (RAG) systems, where users interact with AI assistants through multi-turn conversations containing context-dependent queries. We propose RCEM, a conversational dense retrieval model that distills the query reformulation capability of LLMs into the embedding model, enabling context-aware retrieval without explicit query rewriting during inference. Unlike prior conversational dense retrieval approaches that learn direct conversation-to-document matching, RCEM aligns conversational-query embeddings with rewritten-query embeddings, improving robustness under distributional shift. RCEM does not require conversational query-to-document relevance mappings for training, which are often expensive and difficult to obtain with high quality. Extensive experiments on QReCC, TopiOCQA, and TREC CAsT demonstrate that RCEM consistently outperforms strong conversational retrieval baselines, achieving particularly large gains under distributional shift, including up to 20% improvement in Recall@10. RCEM further extends the base embedding model with conversational query rewriting capability while preserving its original retrieval functionality, allowing both standalone and conversational queries to be encoded by a single model and searched against existing document indexes without rebuilding the retrieval database.

Liu He, Yijuan Lu, John Corring et al.

We develop a diffusion-based approach for various document layout sequence generation. Layout sequences specify the contents of a document design in an explicit format. Our novel diffusion-based approach works in the sequence domain rather than the image domain in order to permit more complex and realistic layouts. We also introduce a new metric, Document Earth Mover's Distance (Doc-EMD). By considering similarity between heterogeneous categories document designs, we handle the shortcomings of prior document metrics that only evaluate the same category of layouts. Our empirical analysis shows that our diffusion-based approach is comparable to or outperforming other previous methods for layout generation across various document datasets. Moreover, our metric is capable of differentiating documents better than previous metrics for specific cases.

Ting-Wu Chin, Cha Zhang, Diana Marculescu

Fine-tuning through knowledge transfer from a pre-trained model on a large-scale dataset is a widely spread approach to effectively build models on small-scale datasets. In this work, we show that a recent adversarial attack designed for transfer learning via re-training the last linear layer can successfully deceive models trained with transfer learning via end-to-end fine-tuning. This raises security concerns for many industrial applications. In contrast, models trained with random initialization without transfer are much more robust to such attacks, although these models often exhibit much lower accuracy. To this end, we propose noisy feature distillation, a new transfer learning method that trains a network from random initialization while achieving clean-data performance competitive with fine-tuning. Code available at https://github.com/cmu-enyac/Renofeation.

Ting-Wu Chin, Ruizhou Ding, Cha Zhang et al.

Pruning convolutional filters has demonstrated its effectiveness in compressing ConvNets. Prior art in filter pruning requires users to specify a target model complexity (e.g., model size or FLOP count) for the resulting architecture. However, determining a target model complexity can be difficult for optimizing various embodied AI applications such as autonomous robots, drones, and user-facing applications. First, both the accuracy and the speed of ConvNets can affect the performance of the application. Second, the performance of the application can be hard to assess without evaluating ConvNets during inference. As a consequence, finding a sweet-spot between the accuracy and speed via filter pruning, which needs to be done in a trial-and-error fashion, can be time-consuming. This work takes a first step toward making this process more efficient by altering the goal of model compression to producing a set of ConvNets with various accuracy and latency trade-offs instead of producing one ConvNet targeting some pre-defined latency constraint. To this end, we propose to learn a global ranking of the filters across different layers of the ConvNet, which is used to obtain a set of ConvNet architectures that have different accuracy/latency trade-offs by pruning the bottom-ranked filters. Our proposed algorithm, LeGR, is shown to be 2x to 3x faster than prior work while having comparable or better performance when targeting seven pruned ResNet-56 with different accuracy/FLOPs profiles on the CIFAR-100 dataset. Additionally, we have evaluated LeGR on ImageNet and Bird-200 with ResNet-50 and MobileNetV2 to demonstrate its effectiveness. Code available at https://github.com/cmu-enyac/LeGR.

Xingyu Fu, Minqian Liu, Zhengyuan Yang et al.

Structured image understanding, such as interpreting tables and charts, requires strategically refocusing across various structures and texts within an image, forming a reasoning sequence to arrive at the final answer. However, current multimodal large language models (LLMs) lack this multihop selective attention capability. In this work, we introduce ReFocus, a simple yet effective framework that equips multimodal LLMs with the ability to generate "visual thoughts" by performing visual editing on the input image through code, shifting and refining their visual focuses. Specifically, ReFocus enables multimodal LLMs to generate Python codes to call tools and modify the input image, sequentially drawing boxes, highlighting sections, and masking out areas, thereby enhancing the visual reasoning process. We experiment upon a wide range of structured image understanding tasks involving tables and charts. ReFocus largely improves performance on all tasks over GPT-4o without visual editing, yielding an average gain of 11.0% on table tasks and 6.8% on chart tasks. We present an in-depth analysis of the effects of different visual edits, and reasons why ReFocus can improve the performance without introducing additional information. Further, we collect a 14k training set using ReFocus, and prove that such visual chain-of-thought with intermediate information offers a better supervision than standard VQA data, reaching a 8.0% average gain over the same model trained with QA pairs and 2.6% over CoT.

Li Sun, Florian Luisier, Kayhan Batmanghelich et al.

Current state-of-the-art models for natural language understanding require a preprocessing step to convert raw text into discrete tokens. This process known as tokenization relies on a pre-built vocabulary of words or sub-word morphemes. This fixed vocabulary limits the model's robustness to spelling errors and its capacity to adapt to new domains. In this work, we introduce a novel open-vocabulary language model that adopts a hierarchical two-level approach: one at the word level and another at the sequence level. Concretely, we design an intra-word module that uses a shallow Transformer architecture to learn word representations from their characters, and a deep inter-word Transformer module that contextualizes each word representation by attending to the entire word sequence. Our model thus directly operates on character sequences with explicit awareness of word boundaries, but without biased sub-word or word-level vocabulary. Experiments on various downstream tasks show that our method outperforms strong baselines. We also demonstrate that our hierarchical model is robust to textual corruption and domain shift.

Baoguang Shi, Wenfeng Cheng, Yijuan Lu et al.

We study the problem of recognizing structured text, i.e. text that follows certain formats, and propose to improve the recognition accuracy of structured text by specifying regular expressions (regexes) for biasing. A biased recognizer recognizes text that matches the specified regexes with significantly improved accuracy, at the cost of a generally small degradation on other text. The biasing is realized by modeling regexes as a Weighted Finite-State Transducer (WFST) and injecting it into the decoder via dynamic replacement. A single hyperparameter controls the biasing strength. The method is useful for recognizing text lines with known formats or containing words from a domain vocabulary. Examples include driver license numbers, drug names in prescriptions, etc. We demonstrate the efficacy of regex biasing on datasets of printed and handwritten structured text and measures its side effects.

Minghao Li, Tengchao Lv, Jingye Chen et al.

Text recognition is a long-standing research problem for document digitalization. Existing approaches are usually built based on CNN for image understanding and RNN for char-level text generation. In addition, another language model is usually needed to improve the overall accuracy as a post-processing step. In this paper, we propose an end-to-end text recognition approach with pre-trained image Transformer and text Transformer models, namely TrOCR, which leverages the Transformer architecture for both image understanding and wordpiece-level text generation. The TrOCR model is simple but effective, and can be pre-trained with large-scale synthetic data and fine-tuned with human-labeled datasets. Experiments show that the TrOCR model outperforms the current state-of-the-art models on the printed, handwritten and scene text recognition tasks. The TrOCR models and code are publicly available at \url{https://aka.ms/trocr}.

Yiheng Xu, Tengchao Lv, Lei Cui et al.

Multimodal pre-training with text, layout, and image has achieved SOTA performance for visually-rich document understanding tasks recently, which demonstrates the great potential for joint learning across different modalities. In this paper, we present LayoutXLM, a multimodal pre-trained model for multilingual document understanding, which aims to bridge the language barriers for visually-rich document understanding. To accurately evaluate LayoutXLM, we also introduce a multilingual form understanding benchmark dataset named XFUND, which includes form understanding samples in 7 languages (Chinese, Japanese, Spanish, French, Italian, German, Portuguese), and key-value pairs are manually labeled for each language. Experiment results show that the LayoutXLM model has significantly outperformed the existing SOTA cross-lingual pre-trained models on the XFUND dataset. The pre-trained LayoutXLM model and the XFUND dataset are publicly available at https://aka.ms/layoutxlm.

Yang Xu, Yiheng Xu, Tengchao Lv et al.

Pre-training of text and layout has proved effective in a variety of visually-rich document understanding tasks due to its effective model architecture and the advantage of large-scale unlabeled scanned/digital-born documents. We propose LayoutLMv2 architecture with new pre-training tasks to model the interaction among text, layout, and image in a single multi-modal framework. Specifically, with a two-stream multi-modal Transformer encoder, LayoutLMv2 uses not only the existing masked visual-language modeling task but also the new text-image alignment and text-image matching tasks, which make it better capture the cross-modality interaction in the pre-training stage. Meanwhile, it also integrates a spatial-aware self-attention mechanism into the Transformer architecture so that the model can fully understand the relative positional relationship among different text blocks. Experiment results show that LayoutLMv2 outperforms LayoutLM by a large margin and achieves new state-of-the-art results on a wide variety of downstream visually-rich document understanding tasks, including FUNSD (0.7895 $\to$ 0.8420), CORD (0.9493 $\to$ 0.9601), SROIE (0.9524 $\to$ 0.9781), Kleister-NDA (0.8340 $\to$ 0.8520), RVL-CDIP (0.9443 $\to$ 0.9564), and DocVQA (0.7295 $\to$ 0.8672). We made our model and code publicly available at \url{https://aka.ms/layoutlmv2}.

Zhengyuan Yang, Yijuan Lu, Jianfeng Wang et al.

In this paper, we propose Text-Aware Pre-training (TAP) for Text-VQA and Text-Caption tasks. These two tasks aim at reading and understanding scene text in images for question answering and image caption generation, respectively. In contrast to the conventional vision-language pre-training that fails to capture scene text and its relationship with the visual and text modalities, TAP explicitly incorporates scene text (generated from OCR engines) in pre-training. With three pre-training tasks, including masked language modeling (MLM), image-text (contrastive) matching (ITM), and relative (spatial) position prediction (RPP), TAP effectively helps the model learn a better aligned representation among the three modalities: text word, visual object, and scene text. Due to this aligned representation learning, even pre-trained on the same downstream task dataset, TAP already boosts the absolute accuracy on the TextVQA dataset by +5.4%, compared with a non-TAP baseline. To further improve the performance, we build a large-scale dataset based on the Conceptual Caption dataset, named OCR-CC, which contains 1.4 million scene text-related image-text pairs. Pre-trained on this OCR-CC dataset, our approach outperforms the state of the art by large margins on multiple tasks, i.e., +8.3% accuracy on TextVQA, +8.6% accuracy on ST-VQA, and +10.2 CIDEr score on TextCaps.

Ross Cutler, Ramin Mehran, Sam Johnson et al.

Active speaker detection (ASD) and virtual cinematography (VC) can significantly improve the remote user experience of a video conference by automatically panning, tilting and zooming of a video conferencing camera: users subjectively rate an expert video cinematographer's video significantly higher than unedited video. We describe a new automated ASD and VC that performs within 0.3 MOS of an expert cinematographer based on subjective ratings with a 1-5 scale. This system uses a 4K wide-FOV camera, a depth camera, and a microphone array; it extracts features from each modality and trains an ASD using an AdaBoost machine learning system that is very efficient and runs in real-time. A VC is similarly trained using machine learning to optimize the subjective quality of the overall experience. To avoid distracting the room participants and reduce switching latency the system has no moving parts -- the VC works by cropping and zooming the 4K wide-FOV video stream. The system was tuned and evaluated using extensive crowdsourcing techniques and evaluated on a dataset with N=100 meetings, each 2-5 minutes in length.

Aaditya Prakash, James Storer, Dinei Florencio et al.

A well-trained Convolutional Neural Network can easily be pruned without significant loss of performance. This is because of unnecessary overlap in the features captured by the network's filters. Innovations in network architecture such as skip/dense connections and Inception units have mitigated this problem to some extent, but these improvements come with increased computation and memory requirements at run-time. We attempt to address this problem from another angle - not by changing the network structure but by altering the training method. We show that by temporarily pruning and then restoring a subset of the model's filters, and repeating this process cyclically, overlap in the learned features is reduced, producing improved generalization. We show that the existing model-pruning criteria are not optimal for selecting filters to prune in this context and introduce inter-filter orthogonality as the ranking criteria to determine under-expressive filters. Our method is applicable both to vanilla convolutional networks and more complex modern architectures, and improves the performance across a variety of tasks, especially when applied to smaller networks.

Ting-Wu Chin, Cha Zhang, Diana Marculescu

Resource-efficient convolution neural networks enable not only the intelligence on edge devices but also opportunities in system-level optimization such as scheduling. In this work, we aim to improve the performance of resource-constrained filter pruning by merging two sub-problems commonly considered, i.e., (i) how many filters to prune for each layer and (ii) which filters to prune given a per-layer pruning budget, into a global filter ranking problem. Our framework entails a novel algorithm, dubbed layer-compensated pruning, where meta-learning is involved to determine better solutions. We show empirically that the proposed algorithm is superior to prior art in both effectiveness and efficiency. Specifically, we reduce the accuracy gap between the pruned and original networks from 0.9% to 0.7% with 8x reduction in time needed for meta-learning, i.e., from 1 hour down to 7 minutes. To this end, we demonstrate the effectiveness of our algorithm using ResNet and MobileNetV2 networks under CIFAR-10, ImageNet, and Bird-200 datasets.

Jingdong Wang, Yajie Xing, Kexin Zhang et al.

Identity transformations, used as skip-connections in residual networks, directly connect convolutional layers close to the input and those close to the output in deep neural networks, improving information flow and thus easing the training. In this paper, we introduce two alternative linear transforms, orthogonal transformation and idempotent transformation. According to the definition and property of orthogonal and idempotent matrices, the product of multiple orthogonal (same idempotent) matrices, used to form linear transformations, is equal to a single orthogonal (idempotent) matrix, resulting in that information flow is improved and the training is eased. One interesting point is that the success essentially stems from feature reuse and gradient reuse in forward and backward propagation for maintaining the information during flow and eliminating the gradient vanishing problem because of the express way through skip-connections. We empirically demonstrate the effectiveness of the proposed two transformations: similar performance in single-branch networks and even superior in multi-branch networks in comparison to identity transformations.

Emad Barsoum, Cha Zhang, Cristian Canton Ferrer et al.

Crowd sourcing has become a widely adopted scheme to collect ground truth labels. However, it is a well-known problem that these labels can be very noisy. In this paper, we demonstrate how to learn a deep convolutional neural network (DCNN) from noisy labels, using facial expression recognition as an example. More specifically, we have 10 taggers to label each input image, and compare four different approaches to utilizing the multiple labels: majority voting, multi-label learning, probabilistic label drawing, and cross-entropy loss. We show that the traditional majority voting scheme does not perform as well as the last two approaches that fully leverage the label distribution. An enhanced FER+ data set with multiple labels for each face image will also be shared with the research community.

Pengfei Wan, Gene Cheung, Philip A. Chou et al.

In texture-plus-depth representation of a 3D scene, depth maps from different camera viewpoints are typically lossily compressed via the classical transform coding / coefficient quantization paradigm. In this paper we propose to reduce distortion of the decoded depth maps due to quantization. The key observation is that depth maps from different viewpoints constitute multiple descriptions (MD) of the same 3D scene. Considering the MD jointly, we perform a POCS-like iterative procedure to project a reconstructed signal from one depth map to the other and back, so that the converged depth maps have higher precision than the original quantized versions.