Tables in business documents power analytics and compliance, yet task-specific datasets are costly to build. Practitioners therefore turn to zero-shot vision–language models (VLMs). We study zero-shot realism for table detection (TD) and table structure recognition (TSR) under a unified protocol on DocILE-QUEST and a private STM154 corpus. We report TD with GIoU, Purity, and Completeness, and TSR with TEDS and TEDS-S, evaluating commercial VLMs (GPT-4o, GPT-5-mini), compact detectors, and supervised YOLO/DETR baselines. Zero-shot VLMs are strong for TSR and competitive for TD, while fine-tuned or from-scratch detectors lead when box quality and robustness to clutter matter. We add an automated error taxonomy that isolates actionable failures (missed, merged/split tables, header–body confusions, cell topology). Finally, we quantify emissions, finding a 104gap between the lightest and heaviest systems.

Document classification systems become more and more complex with the need to deal with new document formats or categories while obtaining a low error rate when classifying more and more documents. Such systems need to have important features including (1) the ability to eliminate ambiguity to improve precision or reduce the error rate, (2) the capability to detect and reject documents belonging to new categories or new variations. Previous studies often focused on closed datasets or solely on the problem of novelty detection, without evaluating the ability to reject ambiguous results after the novelty detection. In this paper, we propose an end-to-end document classification algorithm including both novelty and ambiguity rejection. The proposed algorithm utilizes deep metric learning to compact the knowledge space, and then uses the last hidden layer’s features as input for an unsupervised KNN-based method for novelty and ambiguity rejection. Extensive experiments and analysis on private and public benchmark datasets demonstrate the effectiveness of our proposed algorithm. The algorithm provides the capability to handle new documents while effectively rejecting ambiguity to enhance the precision, recall of known categories, and coverage rate of the end-to-end document classification system.

We propose both a new strategy to weight a text vector for document classification and a comparison of a deep learning approach versus an incremental classification approach, integrating our novel strategy. Bag-of-word vectors are classic approaches to describe a textual document in document classification objective. A weakness of the bag of words is to lose the organization of the discourse within the document. Inspired by some Deep Learning approaches and Natural Language Processing for text classification, we suggest a simple strategy, featuring the terms according to their relative positions within the discourse sequence. For experimentations, we apply this strategy to a recent document incremental classification approach from the state-of-the-art. And, we propose an original comparison between Incremental learning and Deep learning, by comparing the incremental system with a CCN-RNN-based approach. It demonstrates that both approaches are competitive in similar contest.

Our objective is to propose a Document Understanding System for Digital Mailroom application which can cope with three challenges: (1) process a full workflow with high accuracy, with the constraint of a partial training; (2) minimal requirement for configuration work from expert users; (3) adapt incrementally the system in quasi real-time to continuously maximize the recall. We describe an end-to-end system based on existing incremental algorithms for both document classification and field extraction. But in this paper, we really focus on the document classification issue. The main contribution is to adapt the Incremental Growing Neural Gas (A2ING) with a dynamic incremental feature vector. Moreover, a generic Framework automatically selects textual descriptors relying on performance. The quality assessment converges the A2ING and controls the system accuracy.

In this paper, we describe a method to restore a flow of continuous documents. The flow is a collection of consecutive scanned pages without explicit separation marks between documents. Our method is based on contextual and layout descriptors meant to specify the relationship between each pair of consecutive pages. The relationships are represented using vectors of features with boolean values indicating the presence or the absence of descriptors on concerned pages. The segmentation task therefore consists in classifying such vectors into continuities or breaks. The continuity class indicates that pages belong to the same document while the break class ends the ongoing document and starts a new one. The experimental part is based on a large collection of real administrative documents.

This paper proposes an entity local structure comparison approach based on inexact subgraph matching. The comparison results are used for mislabeling correction in the local structure. The latter represents a set of entity attribute labels which are physically close in a document image. It is modeled by an attributed graph describing the content and presentation features of the labels by the nodes and the geometrical features by the arcs. A local structure graph is matched with a structure model which represents a set of local structure model graphs. The structure model is initially built using a set of well chosen local structures based on a graph clustering algorithm and is then incrementally updated. The subgraph matching adopts a specific cost function that integrates the feature dissimilarities. The matched model graph is used to extract the missed labels, prune the extraneous ones and correct the erroneous label fields in the local structure. The evaluation of the structure comparison approach on 525 local structures extracted from 200 business documents achieves about 90% for recall and 95% for precision. The mislabeling correction rates in these local structures vary between 73% and 100%.