Sabine Barrat

Histogram of radon transform. A useful descriptor for shape retrieval

In this paper we present a new descriptor based on the Radon transform. We propose a histogram of the Radon transform, called HRT, which is invariant to common geometrical transformations. For black and white shapes, the HRT descriptor is a histogram of shape lengths at each orientation. The experimental results, defined on different databases and compared with several well-known descriptors, show the robustness of our method.

A Bayesian network for combining descriptors: application to symbol recognition

In this paper, we propose a descriptor combination method, which enables to improve significantly the recognition rate compared to the recognition rates obtained by each descriptor. This approach is based on a probabilistic graphical model. This model also enables to handle both discrete and continuous-valued variables. In fact, in order to improve the recognition rate, we have combined two kinds of features: discrete features (corresponding to shape measures) and continuous features (corresponding to shape descriptors). In order to solve the dimensionality problem due to the large dimension of visual features, we have adapted a variable selection method. Experimental results, obtained in a supervised learning context, on noisy and occluded symbols, show the feasibility of the approach.

A Contour-Based Method for Logo Detection

This paper presents a new approach for logo detection exploiting contour based features. At first stage, pre-processing, contour detection and line segmentation are done. These processes result in set of Outer Contour Strings (OCSs) describing each graphics and text parts of the documents. Then, the logo detection problem is defined as a region scoring problem. Two types of features, coarse and finer ones, are computed from each OCS. Coarse features catch graphical and domain information about OCSs, such as logo positions and aspect ratios. Finer features characterize the contour regions using a gradient based representation. Using these features, we employ regression fitting to score how likely an OCS takes part of a logo region. A final step of correction helps with the wrong segmentation cases. We present experiments done on the Tobacco-800 dataset, and compare our results with the literature. We obtain interesting results compared to the best systems.

Accurate junction detection and characterization in line-drawing images

In this paper, we present a new approach for junction detection and characterization in line-drawing images. We formulate this problem as searching for optimal meeting points of median lines. In this context, the main contribution of the proposed approach is three-fold. First, a new algorithm for the determination of the support region is presented using the linear least squares technique, making it robust to digitization effects. Second, an efficient algorithm is proposed to detect and conceptually remove all distorted zones, retaining reliable line segments only. These line segments are then locally characterized to form a local structure representation of each crossing zone. Finally, a novel optimization algorithm is presented to reconstruct the junctions. Junction characterization is then simply derived. The proposed approach is very highly robust to common geometry transformations and can resist a satisfactory level of noise/degradation. Furthermore, it works very efficiently in terms of time complexity and requires no prior knowledge of the document content. Extensive evaluations have been performed to validate the proposed approach using other baseline methods. An application of symbol spotting is also provided, demonstrating quite good results. HighlightsWe present a new approach for junction detection in line-drawing documents.We present a novel algorithm to deal with the problem of junction distortion.We present an efficient junction optimization algorithm.The characterization of the detected junctions is presented.We obtained very good results relative to the baseline methods.

Classification and Automatic Annotation Extension of Images Using Bayesian Network

In many vision problems, instead of having fully annotated training data, it is easier to obtain just a subset of data with annotations, because it is less restrictive for the user. For this reason, in this paper, we consider especially the problem of classifying weakly-annotated images, where just a small subset of the database is annotated with keywords. In this paper we present and evaluate a new method which improves the effectiveness of content-based image classification, by integrating semantic concepts extracted from text, and by automatically extending annotations to the images with missing keywords. Our model is inspired from the probabilistic graphical model theory: we propose a hierarchical mixture model which enables to handle missing values. Results of visual-textual classification, reported on a database of images collected from the Web, partially and manually annotated, show an improvement by 32.3% in terms of recognition rate against only visual information classification. Besides the automatic annotation extension with our model for images with missing keywords outperforms the visual-textual classification by 6.8%. Finally the proposed method is experimentally competitive with the state-of-art classifiers.

A Robust Approach for Local Interest Point Detection in Line-Drawing Images

In this paper, we propose a new method to detect local interest points as junctions in line-drawing images. Our approach takes advantages of different aspects. Firstly, we extract skeleton of image and then construct a Skeleton Connective Graph with the expectation that it provides a first level of junction detection from shapes. Secondly, instead of employing low-level operators to detect junctions as described in many traditional techniques, our method works at path level taking different skeleton branches into account to gain robustness. Thirdly, we exploit the benefits of wavelet transform (e.g. multi-resolution analysis, discontinuity detection, fast computation, less sensitive to noises) to efficiently detect the dominant points from 1D representations of the paths. Finally, a post-process of pruning and connecting the skeleton segments is performed to discard false detected points and to refine the skeleton. We present in experiments interesting results compared to different methods.

Modeling, classifying and annotating weakly annotated images using Bayesian network

In this paper, we propose a probabilistic graphical model to represent weakly annotated images. We consider an image as weakly annotated if the number of keywords defined for it is less than the maximum number defined in the ground truth. This model is used to classify images and automatically extend existing annotations to new images by taking into account semantic relations between keywords. The proposed method has been evaluated in visual-textual classification and automatic annotation of images. The visual-textual classification is performed by using both visual and textual information. The experimental results, obtained from a database of more than 30,000 images, show an improvement by 50.5% in terms of recognition rate against only visual information classification. Taking into account semantic relations between keywords improves the recognition rate by 10.5%. Moreover, the proposed model can be used to extend existing annotations to weakly annotated images, by computing distributions of missing keywords. Semantic relations improve the mean rate of good annotations by 6.9%. Finally, the proposed method is competitive with a state-of-art model.

An Approximate Proximity Graph Incremental Construction for Large Image Collections Indexing

This paper addresses the problem of the incremental construction of an indexing structure, namely a proximity graph, for large image collections. To this purpose, a local update strategy is examined. Considering an existing graph G and a new node q, how only a relevant sub-graph of G can be updated following the insertion of q? For a given proximity graph, we study the most recent algorithm of the literature and highlight its limitations. Then, a method that leverages an edge-based neighbourhood local update strategy to yield an approximate graph is proposed. Using real-world and synthetic data, the proposed algorithm is tested to assess the accuracy of the approximate graphs. The scalability is verified with large image collections, up to one million images.

A New One-Class Classification Method Based on Symbolic Representation: Application to Document Classification

Training a system using a small number of instances to obtain accurate recognition/classification is a crucial need in document classification domain. The one-class classification is chosen since only positive samples are available for the training. In this paper, a new one-class classification method based on symbolic representation method is proposed. Initially a set of features is extracted from the training set. A set of intervals valued symbolic feature vector is then used to represent the class. Each interval value (symbolic data) is computed using mean and standard deviation of the corresponding feature values. To evaluate the proposed one-class classification method a dataset composed of 544 document images was used. Experiment results reveal that the proposed one-class classification method works well even when the number of training samples is small (≤10). Moreover, we noted that the proposed one-class classification method is suitable for document classification and provides better result compared to one-class k-nearest neighbor (k-NN) classifier.

An Efficient Indexing Scheme Based on Linked-Node m-Ary Tree Structure

Fast nearest neighbor search is a crucial need for many recognition systems. Despite the fact that a large number of indexing algorithms have been proposed in the literature, few of them (e.g., randomized KD-trees, hierarchical K-means tree, randomized clustering trees, and LHS-based schemes) have been well validated on extensive experiments to give satisfactory performance on specific benchmarks. In this work, we propose a linked-node m-ary tree (LM-tree) algorithm, which works really well for both exact and approximate nearest neighbor search. The main contribution of the LM-tree is three-fold. First, a new polar-space-based method of data decomposition is presented to construct the LM-tree. Second, a novel pruning rule is proposed to efficiently narrow down the search space. Finally, a bandwidth search method is introduced to explore the nodes of the LM-tree. Our experiments, applied to one million 128-dimensional SIFT features and 250000 960-dimensional GIST features, showed that the proposed algorithm gives the best search performance, compared to the aforementioned algorithms.