Blind deconvolution is a restoration process of an image which is blurred by an unknown point spread function (PSF) (a.k.a. blur kernel). Some previous works [1], [2], [3], [4] have shown that multiple blurred captures of the same scene can improve the quality of blind deconvolution result. However, the previous multi-image blind deconvolution methods are prone to inconsistencies between observations such as moving objects, illumination change or mis-alignment. In this paper, we present a new multi-image blind deconvolution algorithm using low-rank representation which utilizes similar components of the multiple images for collaboration. Our framework alternatively solves a Schatten-0 norm low-rank approximation and a MAP-based L0norm blind deconvolution for finding the true latent images and their corresponding PSFs and warping parameters. The experimental results show that our approach can recover high quality images in the presence of possible corruptions on both static and moving scenes and outperforms the state-of-the-art results.

We address the problem of detecting and recognizing the text embedded in online images that are circulated over the Web. Our idea is to leverage context information for both text detection and recognition. For detection, we use local image context around the text region, based on that the text often sequentially appear in online images. For recognition, we exploit the metadata associated with the input online image, including tags, comments, and title, which are used as a topic prior for the word candidates in the image. To infuse such two sets of context information, we propose a contextual text spotting network (CTSN). We perform comparative evaluation with five state-of-the-art text spotting methods on newly collected Instagram and Flickr datasets. We show that our approach that benefits from context information is more successful for text spotting in online images.

In this paper, we solve blind image deconvolution problem that is to remove blurs form a signal degraded image without any knowledge of the blur kernel. Since the problem is ill-posed, an image prior plays a significant role in accurate blind deconvolution. Traditional image prior assumes coefficients in filtered domains are sparse. However, it is assumed here that there exist additional structures over the sparse coefficients. Accordingly, we propose new problem formulation for the blind image deconvolution, which utilize the structural information by coupling Student's-t image prior with overlapping group sparsity. The proposed method resulted in an effective blind deconvolution algorithm that outperforms other state-of-the-art algorithms.

When modeling the background with kernel density estimation, the selection of a proper kernel bandwidth becomes a critical issue. It is not easy, however, to perform pixel-wise kernel bandwidth estimation when the data associated with each pixel is insufficient. In this paper, we present a new method using spatial information to estimate the pixel-wise kernel bandwidth. The number of pixels in a spatial region is large enough to capture the variance of the underlying distribution on which the optimal kernel bandwidth is estimated. To show the effectiveness of the estimated kernel bandwidth, the background subtraction using this bandwidth is applied to OLED defect detection and its result is compared to those using the bandwidths obtained from other approaches.

Defect detection is to find unexpected peak regions in an inspection image. Stable Principal Component Pursuit (SPCP) decomposes a given image into three matrices, low-rank, sparsity, and noise which are used for detecting defects. Each of them contains repeated pattern, spatially narrow abnormal elements which are regarded as defects, and small magnitude elements respectively. However, if the noise level of the image is spatially varied, it is hard to separate noise appropriately using naive SPCP. To overcome the difficulty, we propose a novel sliding-window based SPCP algorithm. First, a repeated pattern of each sliding-window is converted to a matrix for SPCP. The noise level based on rank-one approximation is then estimated, and the matrix decomposition is performed. Finally, the sparsity values of all sliding-windows are merged by averaging, and then the averaged term is used for defect detection. The experimental results show that our algorithm outperforms the traditional approaches.

Image super-resolution typically requires prior knowledge to overcome its under-determinedness. While recent generic prior models give impressive results, we note some common artifacts: blurry edges, jagged edges, staircase surfaces, and inconsistent colors. With understanding the underlying causes, we propose an improved generic prior model using isotropic Huber MRFs. In addition, to avoid the oversimplifying problem that generic models often encounter, we suggest applying them only to the structural component of an image, preserving the textural component. Experiments confirm that our framework successfully eliminates those artifacts while preserving textures. Showing competitive results, the GPU implementation allows near real-time application on video.

Document Object Modeling (DOM) is widely used approach for retrieving data from an XML document. If the size of the XML document is very large, however, using the DOM approach for retrieving data from the XML document may suffer from a lack of memory space for building the associated XML tree in the main memory. To alleviate this problem, we propose a method that allows the very large XML document to be split into small XML documents, retrieves data from the XML tree built from each of these small XML documents, and combines the results from all of the n XML trees to generate the final result. With this proposed approach, the memory space and processing time required to retrieve data from the very large XML document using DOM are reduced so that they can be managed by one single general-purpose personal computer.

In content-based image retrieval (CBIR), content of an image can be expressed in terms of different features such as color, texture, shape, or text annotations. Retrieval based on these features can be various by the way how to combine the feature values. Most of the existing approaches assume a linear relationship between different features, and the usefulness of such systems was limited due to the difficulty in representing high-level concepts using low-level features. In this paper, we introduce Neural Network-based Flexible Image Retrieval (NNFIR) system, a human-computer interaction approach to CBIR using Radial Basis Function (RBF) network to combine the values of the heterogeneous features. By using the RBF network, this approach determines nonlinear relationship between features so that more accurate similarity comparison between images can be supported. The experimental results show that the proposed approach has the superior retrieval performance than the existing linear combining method, the rank-based method and the BackPropagation-based method. Although the proposed retrieval model is for CBIR, it can be easily expanded to handle other media types, such as video and audio.

Agent based information retrieval systems have gained great attention, due to the proliferation of readily available but overwhelming information on the Web. We present a neural network based multi agent, especially hierarchically organized information retrieval system. The motivation is based on our experience with an experimental system called Hierarchically Organized Multi-agent Information Retrieval System (HOMIRS) and an analysis of the tradeoff between multi-agent coordination overhead and the benefits of the hierarchical organization of agents. We first introduce our neural network based information retrieval system and then present experimental results with HOMIRS.

In this paper, we propose a multi-agent learning approach to information retrieval on the World Wide Web where each agent collaboratively learns its environment from user's relevance feedback using a neural network mechanism.Our approach makes it possible to discover information sources that will give the desired information, and retrieve that information efficiently and effectively.First, we present a framework for our multi-agent learning approach and introduce a training procedure for capturing knowledge about user's interests and preferences in the information retrieval domain.Secondly, we mathematically analyze the performance of our approach.Finally, to show the utility of our approach, we present the experimental results of our approach and compare them to those obtained by a traditional search service on the World Wide Web.

Content-based image retrieval system retrieves an image from a database using visual information. Among approaches to expressing visual aspects in queries, "query by sketch" is most convenient and expressive. However, the query drawn by the user is typically quite different from the original image and makes it difficult to retrieve the intended image effectively. In this paper, a new method to retrieve the intended image from imperfect queries is presented. The user-drawn query image is converted to a prime edge graph that represents the topological structure of the objects in the image. The experimental results show that the system retrieves the intended image even from a partial or shifted query.

A structural matching method to inspect the two-dimensional visual pattern is proposed. The pattern is represented by a tree where its nodes contain the polygonal-boundary information derived from the pattern image. To inspect whether or not the inspected pattern is defective, the matching process first checks if the tree constructed from the pattern is isomorphic to the one constructed from the standard defect-free. If they are not isomorphic, the process stops and concludes that the pattern is defective. Otherwise, it performs the second matching step using the polygonal-boundary matching function. The suggested method is illustrated for the PCB inspection and shown to have the much lower time complexity than the conventional methods using graph matching.

The authors previously suggested a graph-aided inference browser in which the knowledge engineers may easily verify the knowledge bases by consulting the sequence of inferring steps. In the suggested inference browser, however, the knowledge engineers themselves keep trying to find the error in the knowledge bases by checking if the displayed inferring sequence faces the undesirable situation. This error-detecting process may be slow and not very accurate as the size of the knowledge base becomes very large. The authors present an extended version of the graph-aided inference browser in which the error detecting process is automatically done. As compared to the previous graph-aided inference browser, the presented inference browser involves one additional module, called an automatic verification module, which detects and analyzes the erroneous situation during the inferring process and corrects the associated errors in the knowledge base.

A tool, called HEXPERT, is presented for building expert systems. HEXPERT has been developed by integrating several techniques and ideas known to relax the difficulty of development of expert systems and to enhance their performance. It provides an environment where the knowledge base is formulated from several subknowledge bases each of which is represented using a object-oriented programming technology and processed in any of three inferring directions, the forward, the backward, and the mixed direction. It also provides two browsers. To transfer data between expert systems and database management systems, HEXPERT can communicate to the INGRES database management system.>

A methodology is given for modeling a problem and solving it using the A* algorithm. The heuristic used for A* is mechanically generated from the simplified problem, which is derived by relaxing each of the predicate formulas describing the rules and the goal state of the problem. The generated heuristic satisfies the conditions of admissibility and monotonicity. The methodology is applicable for solving general problems. The overall procedure for this methodology is illustrated by four well-known problems, namely, the eight-puzzle problem, the traveling salesman problem, the robot planning problem, and the consistent labeling problem. The values of the heuristics generated by this procedure are compared to the corresponding values of problem-oriented heuristics reported in the literature.>