Search

Stereo superpixel segmentation aims to obtain the superpixel segmentation results of the left and right views more cooperatively and consistently, rather than simply performing independent segmentation directly. Thus, the correspondence between two views should be reasonably modeled and fully considered. In this paper, we propose a left-right interactive optimization framework for stereo superpixel segmentation. Considering the disparity in stereo image pairs, we first divide the images into paired region and non-paired region, and propose a collaborative optimization scheme to coordinately refine the matched superpixels of the left and right views in an interactive manner. This is, to the best of our knowledge, the first attempt to generate stereo superpixels considering the parallax consistency. Quantitative and qualitative experiments demonstrate that the proposed framework achieves superior performance in terms of consistency and accuracy compared with single-image superpixel segmentation.

Truth discovery methods and rule-based data repairing methods are two classic lines of approaches to improve data quality in the field of database. Truth discovery methods resolve the multi-source conflicts for the same entity by estimating the reliabilities of different source, while rule-based data repairing methods resolve the inconsistencies among different entities using integrity constraints. However, both lines of methods suffer unsatisfactory performances due to the lacking of enough evidence. In this paper, we propose AutoRepair, a novel automatic multi-source data repairing approach to enrich the evidence by taking the advantages of truth discovery and data repairing. We use functional dependency, one of the most common types of constraints, to detect the violations, and use the source reliability as evidence to discover and repair the errors among these violations. At the same time, the repaired results are used to estimate the source reliability. As the source reliability is unknown in advance, we model the process as an iterative framework to ensure better performance. Extensive experiments are conducted on both simulated and real-world datasets. The results clearly demonstrate the advantages of our approach, which outperform both recent truth discovery and rule-based data repairing methods.

Superpixel decomposition methods are widely used in computer vision and image processing applications. By grouping homogeneous pixels, the accuracy can be increased and the decrease of the number of elements to process can drastically reduce the computational burden. For most superpixel methods, a trade-off is computed between 1) color homogeneity, 2) adherence to the image contours and 3) shape regularity of the decomposition. In this paper, we propose a framework that jointly enforces all these aspects and provides accurate and regular Superpixels with Contour Adherence using Linear Path (SCALP). During the decomposition, we propose to consider color features along the linear path between the pixel and the corresponding superpixel barycenter. A contour prior is also used to prevent the crossing of image boundaries when associating a pixel to a superpixel. Finally, in order to improve the decomposition accuracy and the robustness to noise, we propose to integrate the pixel neighborhood information, while preserving the same computational complexity. SCALP is extensively evaluated on standard segmentation dataset, and the obtained results outperform the ones of the state-of-the-art methods. SCALP is also extended for supervoxel decomposition on MRI images., Computer Vision and Image Understanding (CVIU), 2018

Summary Early quantitative structure-activity relationship (QSAR) technologies have unsatisfactory versatility and accuracy in fields such as drug discovery because they are based on traditional machine learning and interpretive expert features. The development of Big Data and deep learning technologies significantly improve the processing of unstructured data and unleash the great potential of QSAR. Here we discuss the integration of wet experiments (which provide experimental data and reliable verification), molecular dynamics simulation (which provides mechanistic interpretation at the atomic/molecular levels), and machine learning (including deep learning) techniques to improve QSAR models. We first review the history of traditional QSAR and point out its problems. We then propose a better QSAR model characterized by a new iterative framework to integrate machine learning with disparate data input. Finally, we discuss the application of QSAR and machine learning to many practical research fields, including drug development and clinical trials., Graphical abstract, Data analysis in structural biology; Machine learning; Structural biology

Feature selection is an important problem in machine learning and data mining. In reality, the wrapper methods are broadly used in feature selection. It treats feature selection as a search problem using a predictor as a black-box. However, most wrapper methods are time-consuming due to the large search space. In this paper, we propose a novel wrapper method, called FeatureBand, for feature selection. We use the early stopping strategy to terminate bad candidate feature subsets and avoid wasting of training time. Further, we use a genetic local search to generate new subsets based on previous ones. These two techniques are combined under an iterative framework in which we gradually allocate more resources for more promising candidate feature subsets. The experimental result shows that FeatureBand achieves a better trade-off between search time and search accuracy. It is 1.45\(\times \) to 17.6\(\times \) faster than the state-of-the-art wrapper-based methods without accuracy loss.

Nowadays, more and more people have their own accounts in different social networks, and they might use the different email addresses or phone numbers in different networks, so how to identify the same person among different social networks become a vital problem, called network alignment. Users with different accounts are called anchor users, researches showed that using some known anchor users to predict the potential anchor links for the full network is an effective way. To predict more accurate anchor links, the paper proposes a new prediction framework ISS, based on a reality of partially aligned social networks, it applies supervised learning based on social feature extraction and strict stable matching, which improve the accuracy of the prediction result, what is more, we apply an iterative framework to refine known information and maximize the prediction results. Experiments have conducted in two realworld heterogeneous social networks, Foursquare and Twitter, and it demonstrates that ISS can predict anchor links among heterogeneous social networks very well and outperform other similar prediction methods.

Spectral feature selection (SFS) is getting more and more attention in recent years. However, conventional SFS has some weaknesses that may corrupt the performance of feature selection, since (1) SFS generally preserves the either global structure or local structure, which can’t provide comprehensive information for the model; (2) graph learning and feature selection of SFS is two individual processes, which is hard to achieve the global optimization. Thus, a novel SFS is proposed via introducing a low-rank constraint for capturing inherent structure of data, and utilizing an adaptive graph learning to couple the graph learning and feature data learning in an iterative framework to output a robust and accurate learning model. A optimization algorithm is proposed to solve the proposed problem with a fast convergence. By comparing to some classical and first-class feature selection methods, our method has exhibited a competitive performance.

To deal with the rigid template matching problem in real-world scenarios, we propose a novel iterative feature-pair updating framework which is also robust to high levels of outliers, such as background changing, complex nonrigid deformation and partial occlusion. Given a pair of template image and target image, we first extract a set of corresponding feature-pairs as candidates. Then, we propose a robust objective function under the iterative framework for discriminatively updating these candidates, where the space distance, appearance distance, and the overlapping percentage of feature pairs are integrated simultaneously. Finally, a hierarchical matching strategy is provided with the parameter discussion. Experimental results compared with the-state-of-art methods on public data sets demonstrate the effectiveness of the proposed method.

The goal of an object category discovery system is to annotate a pool of unlabeled image data, where the set of labels is initially unknown to the system, and must therefore be discovered over time by querying a human annotator. The annotated data is then used to train object detectors in a standard supervised learning setting, possibly in conjunction with category discovery itself. Category discovery systems can be evaluated in terms of both accuracy of the resulting object detectors, and the efficiency with which they discover categories and annotate the training data. To improve the accuracy and efficiency of category discovery, we propose an iterative framework which alternates between optimizing nearest neighbor classification for known categories with multiple kernel metric learning, and detecting clusters of unlabeled image regions likely to belong to a novel, unknown categories. Experimental results on the MSRC and PASCAL VOC2007 data sets show that the proposed method improves clustering for category discovery, and efficiently annotates image regions belonging to the discovered classes.

Reassembling hand shredded content-less pages is a challenging task, with applications in forensics and fun games. This paper proposes an efficient iterative framework to solve apictorial jigsaw puzzles of hand shredded content-less pages, using only the shape information. The proposed framework consists of four phases. In the first phase, normalized shape features are extracted from fragment contours. Then, for all possible matches between pairs of fragments transformation parameters for alignment of fragments and three goodness scores are estimated. In the third phase, incorrect matches are eliminated based on the score values. The alignments are refined by pruning the set of pairwise matched fragments. Finally, a modified graph-based framework for agglomerative clustering is used to globally reassemble the page(s). Experimental evaluation of our proposed framework on an annotated dataset of shredded documents shows the efficiency in the reconstruction of multiple content-less pages from arbitrarily torn fragments.

Hand shredded content-less pages reassembly is a challenging task. This has applications in forensics and fun games. The process is even more tedious when the number of pages from which the fragments are obtained is unknown. An iterative framework to solve the jigsaw puzzles of multiple hand shredded content-less pages has been proposed in this paper. This framework makes use of the shape-based information alone to solve the puzzle. All pairs of fragments are matched using the normalized shape-based features. Then, incorrect matches between the fragments are pruned using three scores that measure the goodness of the alignment. Finally, a graph-based technique is used to densely arrange the fragments for the global reassembly of the page(s). Experimental evaluation of our proposed framework on an annotated dataset of shredded documents shows the efficiency in the reconstruction of multiple content-less pages from arbitrarily torn fragments and performance metrics have been proposed to numerically evaluate the reassembly.

Spatial co-location pattern mining discovers the subsets of features whose events are frequently located together in geographic space. The current research on this topic adopts a threshold-based approach that requires users to specify in advance the thresholds of distance and prevalence. However, in practice, it is not easy to specify suitable thresholds. In this article, we propose a novel iterative mining framework that discovers spatial co-location patterns without predefined thresholds. With the absolute and relative prevalence of spatial co-locations, our method allows users to iteratively select informative edges to construct the neighborhood relationship graph until every significant co-location has enough confidence and eventually to discover all spatial co-location patterns. The experimental results on real world data sets indicate that our framework is effective for prevalent co-locations discovery.

In this paper, we consider a speech reconstruction problem, which is efficiently solved via sparse representation. Though a variety of speech reconstruction methods based on the sparse representation are developed, they seldom take into account the intrinsic attributes of speech, e.g., harmonic structures. To address this issue, a harmonic-based sparse representation algorithm is proposed to emphasize harmonic correlations between the adjacent speech frames. Sequently, a corresponding sparse optimal model is presented with a harmonic regularization term, which can be efficiently solved in an iterative framework. Simulation results demonstrate that the proposed method can achieve a better performance to improve the quality of the reconstructed speech than other traditional sparse representation methods.

This paper presents a novel approach for detection and estimation of fundamental parameters of linear frequency modulation (LFM) signals, i.e., the initial frequency and Chirp rate. The proposed approach is based on sparse representation of noisy input signals over two specific dictionaries, each designed for finding a parameter of LFM signal. Moreover, an iterative framework is proposed for simultaneous sparse representation over the two dictionaries. Experimental results demonstrate that the presented method is comparable with the optimum transform for LFM signal estimation, Wigner-Hough, and furthermore, has significantly higher speed.

International audience; We propose a strategy for semi-supervised learning of Hidden-state Conditional Random Fields (HCRF) for signal classification. It builds on simple procedures for semi-supervised learning of Hidden Markov Models (HMM) and on strategies for learning a HCRF from a trained HMM system. The algorithm learns a generative system based on Hidden Markov models and a discriminative one based on HCRFs where each model is refined by the other in an iterative framework.

This paper presents an algorithm for the segmentation of lung fields by fusing shape information priors into intensity-based thresholding in an iterative framework. The main contribution is to maximize information utilization by effectively combining intensity information with shape priors. Experimental results performed on publicly available database demonstrate the effectiveness of the algorithm in comparison with other algorithms.

Human facial expression modeling and synthesis have recently become a very active area of research. This is partially due to its potential application in model-based image coding, as well as the possibility of using it to enhance human-computer interactions. The majority of the research that has been done in this area has focused on facial expression analysis, modeling and synthesis. Although good results have been obtained in analysis and synthesis, not much effort has been spent on attempting to synthesize facial images that are natural looking. In this paper, we describe our research on facial expression modeling and synthesis. We propose an iterative framework that uses a genetic algorithm to synthesize natural looking facial images. Facial expression representation and distortion measures will also be discussed. Preliminary results are presented.