Tables, especially when having complex layouts, contain rich semantic information. However, effectively learning from tables to uncover such semantic information remains challenging. The rapid progress in natural language processing does not necessarily correspond to equivalent advancements in table parsing, which often requires joint visual and language modeling. Indeed, humans can quickly derive semantic meaning from table entries by associating them with corresponding column and/or row headers. Motivated by this observation, we propose a new heterogeneous Graph-based Table Representation Learning (GTRL) framework. GTRL combines graph-based visual modeling with sequence-based language modeling to learn granular per-cell embeddings that are sensitive to the semantic meaning of cells within their corresponding table context. We systematically evaluate the proposed GTRL framework using two datasets: a new adhesive table benchmark comprising complex tables extracted from industrial documents for learning per-entry semantics, and a publicly available large-scale dataset that enables learning header semantics from column tables. Experimental results demonstrate the competitive performance of the proposed GTRL, which often exhibits reduced computational complexity compared to state-of-the-art table representation learning models.

In this paper we propose a novel probabilistic framework for document segmentation exploiting human perceptual recognition of text regions from complicated layouts. In particular, we conceptualize text homogeneity as the Gestalt pattern displayed in text regions, characterized by proximately and symmetrically arranged units with similar morphological and texture features. We model this pattern in the local region of a connected component (CC) using an hierarchical formulation, which simulates a random walk-and-check on a graph encoding the neighborhood of the CC. The proposed formulation allows an effective computation of what we call the probabilistic local text homogeneity (PLTH) using a weighted summation of the weights of the graph, which are derived from a probabilistic description of the homogeneity between neighboring CCs and computed through Bayesian cue integration. The proposed PLTH enables a multi-aspect analysis, where various primitives such as geometrical configuration, morphological features, texture characterization and location priors are integrated in one computational probabilistic model. This enables an effective text and non-text classification of CCs preceding any grouping process, which is currently absent in document segmentation. Experimental results show that our segmentation method based on the proposed PLTH model improves upon the state-of-the-art.

Physical damages (such as torn-offs and scratches) are commonly seen in historical documents. Recognition of such damages is currently absent in digitization-and-information-extraction (DIE) systems but crucial for automatic document comprehension and exploitation. In this paper we propose a generic damage recognition (DR) method based on a joint global and local modeling of the text homogeneity (TH) pattern exhibited in document images. More specifically, a connected component (CC) based formulation is developed as a global homogeneity measure, where TH is characterized using a probabilistic graph model for a coarse recognition of damaged regions. A multi-resolution analysis (MRA) of TH is further developed for a granular within-CC recognition of damage pixels, where the disparity between damage and text pixels is characterized by exploiting neighborhood transitions. This enables the formulation of a local homogeneity measure, where the neighborhood transition around an individual pixel is modeled using the propagation of the approximation coefficients of a stationary wavelet transform (SWT). The proposed global and local homogeneity measures are integrated as a joint likelihood in a Bayesian model with a Markov random field (MRF) prior, where DR is formulated as a maximum a posterior (MAP) inference which is addressed using Markov Chain Monte Carlo (MCMC) sampling. The resulting algorithm is tested on a set of real-life historical newspaper images containing damages of varying size and shape. The performance of the algorithm is evaluated using both F-measures and the Intersection-over-Union (IoU) metric, where test results demonstrate the promising potential of the proposed method.

Document image quality assessment (DIQA) is an important process for various applications such as optical character recognition (OCR) and document restoration. In this paper we propose a no-reference DIQA model based on a deep convolutional neural network (DCNN), where the rich knowledge of natural scene image characterization of a previously-trained DCNN is exploited towards OCR accuracy oriented document image quality assessment. Following a two-stage deep transfer learning procedure, we fine-tune the knowledge base of the DCNN in the first phase and bring in a task-specific segment consisting of three fully connected (FC) layers in the second phase. Based on the fine-tuned knowledge base, the task-specific segment is trained from scratch to facilitate the application of the transferred knowledge on the new task of document quality assessment. Testing results on a benchmark dataset demonstrate that the proposed model achieves state-of-the-art performance.

Energy consumption represents a significant cost in data center operation. A large fraction of the energy, however, is used to power idle servers when the workload is low. Dynamic provisioning techniques aim at saving this portion of the energy, by turning off unnecessary servers. In this paper, we explore how much gain knowing future workload information can bring to dynamic provisioning. In particular, we develop online dynamic provisioning solutions with and without future workload information available. We first reveal an elegant structure of the offline dynamic provisioning problem, which allows us to characterize the optimal solution in a “divide-andconquer” manner. We then exploit this insight to design two online algorithms with competitive ratios 2 - α and e/(e - 1 + α), respectively, where 0 ≤ α ≤ 1 is the normalized size of a look-ahead window in which future workload information is available. A fundamental observation is that future workload information beyond the full-size look-ahead window (corresponding to α = 1) will not improve dynamic provisioning performance. Our algorithms are decentralized and easy to implement. We demonstrate their effectiveness in simulations using real-world traces.