Search

Document image classification is one of the most important components in business automation workflow. Therefore, a range of different supervised image classification methods have been proposed, which rely on a large amount of labeled data, which is rarely available in practice. Furthermore, retraining of these models is necessary upon the introduction of new classes. In this paper, we analyze the potential of zero-shot document image classification based on computing the agreement between the images and the textual embeddings of class names/descriptions. This enables the deployment of document image classification models without the availability of any training data at zero training cost, alongside providing seamless integration of new classes. Our results show that using zero-shot recognition achieves significantly better than chance performance on document image classification benchmarks (49.51% accuracy on Tobacco-3482 in contrast to 10% random classifier accuracy and 39.22% on RVL-CDIP dataset in contrast to 6.25% random classifier accuracy). We also show that the representation learned by a vision transformer using image-text pairs is competitive to CNNs by training a linear SVM on top of the pre-computed representations which achieves comparable performance to state-of-the-art convolutional networks (85.74% on Tobacco-3482 dataset in contrast to 71.67% from ImageNet pretrained ResNet-50). Even though the initial results look encouraging, there is still a large gap to cover for zero-shot recognition within the domain of document images in contrast to natural scene image classification, which achieves comparable performance to fully supervised baselines. Our preliminary findings pave the way for the deployment of zero-shot document image classification in production settings.

Compression of neural networks has become a common norm in industrial settings to reduce the cost of inference and deployment. As document classification is a common process in business workflows, there is a dire need of analyzing the potential of compressed models for the task of document image classification. Surprisingly, no such analysis has been done in the past. Furthermore, once a compressed model is obtained using a particular compression algorithm (which achieves similar accuracy to the uncompressed model), the model is directly deployed without much consideration. However, such compression process results in its own set of challenges which have large implications, especially when the uncompressed model is ensured to be fair or unbiased such as a fair model making critical hiring decisions based on the applicant’s resume. In this paper, we show that current state-of-the-art compression algorithms can be successfully applied for the task of document image classification. We further analyze the impact of model compression on network outputs and highlight the discrepancy that arises during the compression process. Building on recent findings in this direction, we employ a principled approach based on logit-pairing for minimizing this deviation from the functional outputs of the uncompressed model. Our results on Tobacco-3482 dataset show that we can reduce the number of mismatches between the compressed and the uncompressed model drastically in both structured and unstructured cases (by more than 2x in some cases) by employing label-preservation-aware loss functions. These findings shed light on some important considerations before the compressed models are ready for deployment in the real world.

Traditional distance and density-based anomaly detection techniques are unable to detect periodic and seasonality related point anomalies which occur commonly in streaming data, leaving a big gap in time series anomaly detection in the current era of the IoT. To address this problem, we present a novel deep learning-based anomaly detection approach (DeepAnT) for time series data, which is equally applicable to the non-streaming cases. DeepAnT is capable of detecting a wide range of anomalies, i.e., point anomalies, contextual anomalies, and discords in time series data. In contrast to the anomaly detection methods where anomalies are learned, DeepAnT uses unlabeled data to capture and learn the data distribution that is used to forecast the normal behavior of a time series. DeepAnT consists of two modules: time series predictor and anomaly detector. The time series predictor module uses deep convolutional neural network (CNN) to predict the next time stamp on the defined horizon. This module takes a window of time series (used as a context) and attempts to predict the next time stamp. The predicted value is then passed to the anomaly detector module, which is responsible for tagging the corresponding time stamp as normal or abnormal. DeepAnT can be trained even without removing the anomalies from the given data set. Generally, in deep learning-based approaches, a lot of data are required to train a model. Whereas in DeepAnT, a model can be trained on relatively small data set while achieving good generalization capabilities due to the effective parameter sharing of the CNN. As the anomaly detection in DeepAnT is unsupervised, it does not rely on anomaly labels at the time of model generation. Therefore, this approach can be directly applied to real-life scenarios where it is practically impossible to label a big stream of data coming from heterogeneous sensors comprising of both normal as well as anomalous points. We have performed a detailed evaluation of 15 algorithms on 10 anomaly detection benchmarks, which contain a total of 433 real and synthetic time series. Experiments show that DeepAnT outperforms the state-of-the-art anomaly detection methods in most of the cases, while performing on par with others.

Book covers usually provide a good depiction of a book’s content and its central idea. The classification of books in their respective genre usually involves subjectivity and contextuality. Book retrieval systems would utterly benefit from an automated framework that is able to classify a book’s genre based on an image, specifically for archival documents where digitization of the complete book for the purpose of indexing is an expensive task. While various modalities are available (e.g., cover, title, author, abstract), benchmarking the image-based classification systems based on minimal information is a particularly exciting field due to the recent advancements in the domain of image-based deep learning and its applicability. For that purpose, a natural question arises regarding the plausibility of solving the problem of book classification by only utilizing an image of its cover along with the current state-of-the-art deep learning models. To answer this question, this paper makes a three-fold contribution. First, the publicly available book cover dataset comprising of 57k book covers belonging to 30 different categories is thoroughly analyzed and corrected. Second, it benchmarks the performance on a battery of state-of-the-art image classification models for the task of book cover classification. Third, it uses explicit attention mechanisms to identify the regions that the network focused on in order to make the prediction. All of our evaluations were performed on a subset of the mentioned public book cover dataset. Analysis of the results revealed the inefficacy of the most powerful models for solving the classification task. With the obtained results, it is evident that significant efforts need to be devoted in order to solve this image-based classification task to a satisfactory level.

With the rise in the employment of deep learning methods in safety-critical scenarios, interpretability is more essential than ever before. Although many different directions regarding interpretability have been explored for visual modalities, time series data has been neglected, with only a handful of methods tested due to their poor intelligibility. We approach the problem of interpretability in a novel way by proposing TSInsight, where we attach an auto-encoder to the classifier with a sparsity-inducing norm on its output and fine-tune it based on the gradients from the classifier and a reconstruction penalty. TSInsight learns to preserve features that are important for prediction by the classifier and suppresses those that are irrelevant, i.e., serves as a feature attribution method to boost the interpretability. In contrast to most other attribution frameworks, TSInsight is capable of generating both instance-based and model-based explanations. We evaluated TSInsight along with nine other commonly used attribution methods on eight different time series datasets to validate its efficacy. The evaluation results show that TSInsight naturally achieves output space contraction, therefore, it is an effective tool for the interpretability of deep time series models.

There is a need for automatic systems that can reliably detect, track and classify fish and other marine species in underwater videos without human intervention. Conventional computer vision techniques do not perform well in underwater conditions where the background is complex and the shape and textural features of fish are subtle. Data-driven classification models like neural networks require a huge amount of labelled data, otherwise they tend to over-fit to the training data and fail on unseen test data which is not involved in training. We present a state-of-the-art computer vision method for fine-grained fish species classification based on deep learning techniques. A cross-layer pooling algorithm using a pre-trained Convolutional Neural Network as a generalized feature detector is proposed, thus avoiding the need for a large amount of training data. Classification on test data is performed by a SVM on the features computed through the proposed method, resulting in classification accuracy of 94.3% for fish species from typical underwater video imagery captured off the coast of Western Australia. This research advocates that the development of automated classification systems which can identify fish from underwater video imagery is feasible and a cost-effective alternative to manual identification by humans.

The adversarial vulnerability of deep networks has spurred the interest of researchers worldwide. Unsurprisingly, like images, adversarial examples also translate to time-series data as they are an inherent weakness of the model itself rather than the modality. Several attempts have been made to defend against these adversarial attacks, particularly for the visual modality. In this paper, we perform detailed benchmarking of well-proven adversarial defense methodologies on time-series data. We restrict ourselves to the \(L_{\infty }\) threat model. We also explore the trade-off between smoothness and clean accuracy for regularization-based defenses to better understand the trade-offs that they offer. Our analysis shows that the explored adversarial defenses offer robustness against both strong white-box as well as black-box attacks. This paves the way for future research in the direction of adversarial attacks and defenses, particularly for time-series data.

Identification of input data points relevant for the classifier (i.e. serve as the support vector) has recently spurred the interest of researchers for both interpretability as well as dataset debugging. This paper presents an in-depth analysis of the methods which attempt to identify the influence of these data points on the resulting classifier. To quantify the quality of the influence, we curated a set of experiments where we debugged and pruned the dataset based on the influence information obtained from different methods. To do so, we provided the classifier with mislabeled examples that hampered the overall performance. Since the classifier is a combination of both the data and the model, therefore, it is essential to also analyze these influences for the interpretability of deep learning models. Analysis of the results shows that some interpretability methods can detect mislabels better than using a random approach, however, contrary to the claim of these methods, the sample selection based on the training loss showed a superior performance., Comment: 8 pages, 11 figures, Accepted for the IEEE International Joint Conference on Neural Networks (IJCNN) 2020

This paper presents a novel method for the analysis of tabular structures in document images using the potential of deformable convolutional networks. In order to assess the suitability of the model to the task of table structure recognition, most of the prior methods have been tested on the smaller ICDAR-13 table structure recognition dataset comprising of just 156 tables. We curated a new image-based table structure recognition dataset, TabStructDB2, comprising of 1081 tables densely labeled with row and column information. Instead of collecting new images for this purpose, we leveraged the famous Page-Object Detection dataset from ICDAR-17, and added structural information for all the tabular regions present in the dataset. This new publicly available dataset will enable the development of more sophisticated table structure recognition techniques in the future. We performed extensive evaluation on the two datasets (ICDAR-13 and TabStructDB) including crossdataset testing in order to evaluate the efficacy of the proposed approach. We achieved state-of-the-art results with deformable models on ICDAR-13 with an average F-Measure of 92.98% (89.42% for rows and 96.55% for columns) and report baseline results on TabStructDB for guiding future research efforts with an F-Measure of 93.72% (91.26% for rows and 95.59% for columns). Despite promising results, structural analysis of tables with arbitrary layouts is still far from achievable at this point.

Following publication of the original article

The need for robust unsupervised anomaly detection in streaming data is increasing rapidly in the current era of smart devices, where enormous data are gathered from numerous sensors. These sensors record the internal state of a machine, the external environment, and the interaction of machines with other machines and humans. It is of prime importance to leverage this information in order to minimize downtime of machines, or even avoid downtime completely by constant monitoring. Since each device generates a different type of streaming data, it is normally the case that a specific kind of anomaly detection technique performs better than the others depending on the data type. For some types of data and use-cases, statistical anomaly detection techniques work better, whereas for others, deep learning-based techniques are preferred. In this paper, we present a novel anomaly detection technique, FuseAD, which takes advantage of both statistical and deep-learning-based approaches by fusing them together in a residual fashion. The obtained results show an increase in area under the curve (AUC) as compared to state-of-the-art anomaly detection methods when FuseAD is tested on a publicly available dataset (Yahoo Webscope benchmark). The obtained results advocate that this fusion-based technique can obtain the best of both worlds by combining their strengths and complementing their weaknesses. We also perform an ablation study to quantify the contribution of the individual components in FuseAD, i.e., the statistical ARIMA model as well as the deep-learning-based convolutional neural network (CNN) model.

We propose a novel hybrid approach that fuses traditional computer vision techniques with deep learning models to detect figures and formulas from document images. The proposed approach first fuses the different computer vision based image representations, i.e., color transform, connected component analysis, and distance transform, termed as Fi-Fo image representation. The Fi-Fo image representation is then fed to deep models for further refined representation-learning for detecting figures and formulas from document images. The proposed approach is evaluated on a publicly available ICDAR-2017 Page Object Detection (POD) dataset and its corrected version. It produces the state-of-the-art results for formula and figure detection in document images with an f1-score of 0.954 and 0.922, respectively. Ablation study results reveal that the Fi-Fo image representation helps in achieving superior performance in comparison to raw image representation. Results also establish that the hybrid approach helps deep models to learn more discriminating and refined features.

Neural networks (NN) are considered as black-boxes due to the lack of explainability and transparency of their decisions. This significantly hampers their deployment in environments where explainability is essential along with the accuracy of the system. Recently, significant efforts have been made for the interpretability of these deep networks with the aim to open up the black-box. However, most of these approaches are specifically developed for visual modalities. In addition, the interpretations provided by these systems require expert knowledge and understanding for intelligibility. This indicates a vital gap between the explainability provided by the systems and the novice user. To bridge this gap, we present a novel framework i.e. Time-Series eXplanation (TSXplain) system which produces a natural language based explanation of the decision taken by a NN. It uses the extracted statistical features to describe the decision of a NN, merging the deep learning world with that of statistics. The two-level explanation provides ample description of the decision made by the network to aid an expert as well as a novice user alike. Our survey and reliability assessment test confirm that the generated explanations are meaningful and correct. We believe that generating natural language based descriptions of the network's decisions is a big step towards opening up the black-box.

Propensity of skin diseases to manifest in a variety of forms, lack and maldistribution of qualified dermatologists, and exigency of timely and accurate diagnosis call for automated Computer-Aided Diagnosis (CAD). This study aims at extending previous works on CAD for dermatology by exploring the potential of Deep Learning to classify hundreds of skin diseases, improving classification performance, and utilizing disease taxonomy. We trained state-of-the-art Deep Neural Networks on two of the largest publicly available skin image datasets, namely DermNet and ISIC Archive, and also leveraged disease taxonomy, where available, to improve classification performance of these models. On DermNet we establish new state-of-the-art with 80% accuracy and 98% Area Under the Curve (AUC) for classification of 23 diseases. We also set precedence for classifying all 622 unique sub-classes in this dataset and achieved 67% accuracy and 98% AUC. On ISIC Archive we classified all 7 diseases with 93% average accuracy and 99% AUC. This study shows that Deep Learning has great potential to classify a vast array of skin diseases with near-human accuracy and far better reproducibility. It can have a promising role in practical real-time skin disease diagnosis by assisting physicians in large-scale screening using clinical or dermoscopic images.

This paper presents a novel framework for demystification of convolutional deep learning models for time-series analysis. This is a step towards making informed/explainable decisions in the domain of time-series, powered by deep learning. There have been numerous efforts to increase the interpretability of image-centric deep neural network models, where the learned features are more intuitive to visualize. Visualization in time-series domain is much more complicated as there is no direct interpretation of the filters and inputs as compared to the image modality. In addition, little or no concentration has been devoted for the development of such tools in the domain of time-series in the past. TSViz provides possibilities to explore and analyze a network from different dimensions at different levels of abstraction which includes identification of parts of the input that were responsible for a prediction (including per filter saliency), importance of different filters present in the network for a particular prediction, notion of diversity present in the network through filter clustering, understanding of the main sources of variation learnt by the network through inverse optimization, and analysis of the network's robustness against adversarial noise. As a sanity check for the computed influence values, we demonstrate results regarding pruning of neural networks based on the computed influence information. These representations allow to understand the network features so that the acceptability of deep networks for time-series data can be enhanced. This is extremely important in domains like finance, industry 4.0, self-driving cars, health-care, counter-terrorism etc., where reasons for reaching a particular prediction are equally important as the prediction itself. We assess the proposed framework for interpretability with a set of desirable properties essential for any method., 7 Pages (6 + 1 for references), 7 figures