Anjin Park

Neural Network Implementation Using CUDA and OpenMP

Many algorithms for image processing and pattern recognition have recently been implemented on GPU (graphic processing unit) for faster computational times. However, the implementation using GPU encounters two problems. First, the programmer should master the fundamentals of the graphics shading languages that require the prior knowledge on computer graphics. Second, in a job which needs much cooperation between CPU and GPU, which is usual in image processings and pattern recognitions contrary to the graphics area, CPU should generate raw feature data for GPU processing as much as possible to effectively utilize GPU performance. This paper proposes more quick and efficient implementation of neural networks on both GPU and multi-core CPU. We use CUDA (compute unified device architecture) that can be easily programmed due to its simple C language-like style instead of GPU to solve the first problem. Moreover, OpenMP (Open Multi-Processing) is used to concurrently process multiple data with single instruction on multi-core CPU, which results ineffectively utilizing the memories of GPU. In the experiments, we implemented neural networks-based text detection system using the proposed architecture, and the computational times showed about 15 times faster than implementation using CPU and about 4 times faster than implementation on only GPU without OpenMP.

Automatic conversion system for mobile cartoon contents

As the production of mobile contents is increasing and many people are using it, the existing mobile contents providers manually split cartoons into frame images fitted to the screen of mobile devices. It needs much time and is very expensive. This paper proposes an Automatic Conversion System (ACS) for mobile cartoon contents. It converts automatically the existing cartoon contents into mobile cartoon contents using an image processing technology as follows: 1) A scanned cartoon image is segmented into frames by structure layout analysis. 2) The frames are split at the region that does not include the semantic structure of the original image 3) Texts are extracted from the splitting frames, and located at the bottom of the screen. Our experiment shows that the proposed ACS is more efficient than the existing methods in providing mobile cartoon contents.

Graph Cuts-Based Automatic Color Image Segmentation

A graph cuts method has recently attracted a lot of attention for image segmentation, as it can minimize an energy function composed of data term estimated in feature space and smoothness term estimated in an image domain. Although previous approaches using graph cuts have shown good performance for image segmentation, they manually obtained prior information to estimate the data term, thus automatic image segmentation is one of issues in application using the graph cuts method. To automatically estimate the data term, GMM (Gaussian mixture model) is generally used, but it is practicable only for classes with a hyper-spherical or hyper-ellipsoidal shape, as the class was represented based on the covariance matrix centered on the mean. For arbitrary-shaped classes, this paper proposes graph cuts-based image segmentation using mean shift analysis. As prior information to estimate the data term, we use the set of mean trajectories toward each mode from initial means randomly selected in L*u*v* feature space. Since the mean shift procedure requires many computational times, we transform features in continuous feature space into 3D discrete grid, and use 3D kernel based on the first moment in the grid, which are needed to move the means to modes. In the experiments, we investigated problems of normalized cuts-based and mean shift-based segmentation and graph cuts-based segmentation using GMM. As a result, the proposed method showed better performance than previous three methods on Berkeley segmentation dataset.

Automatic word detection system for document image using mobile devices

In the current age of ubiquitous computing age that uses high bandwidth network, wearable and hand-held mobile devices with small cameras and wireless communication will be widespread in the near future. Thus, computer vision and image processing for mobile devices have recently attracted a lot of attention. Especially, many approaches to detect image texts containing useful information for automatic annotation, indexing, and structuring of image are important for a prerequisite stage of recognition in dictionary application using mobile devices equipped with a camera. To detect image texts on the mobile devices that have limited computational resources, recent works are based on two methodologies; the image texts are detected not by automatically but by manually using stylus pen to reduce the computational resources, and the server is used to detect image texts requiring many floating-computations. The main disadvantage of the manual method is that users directly select tentative text regions, and recall and precision rates are determined by the selected regions. The second method to automatically detect the image texts is difficult to perform it in real-time, due to transmission time between the mobile device and the server. Accordingly, this paper proposes a real-time automatic word detection system without support of the server. To minimize the computational time, one word in the central region of the image is considered as a target of the system. The word region is tentatively extracted by using edge density and window transition, and the tentatively extracted region is then verified by measuring uniform distribution among sub-windows of the extracted region. In the experiments, the proposed method showed high precision rates for one word in the central region of the image, and showed fast computational time on the mobile devices.

Contents recycling using content-based image retrieval on mobile device

Although a lot of studies have been made on mobile learning, the study of content-based image recycling on mobile device is not known very well. This paper presents a new approach which recycles and augments existing off-line contents using a camera-equipped mobile device. Each individual learner has a PDA and an off-line textbook (Picture English Book: PEB). During the PEB-watching learning activity, users are dynamically provided with on-line information such as texts, videos and audios corresponding to the off-line contents via the PDA. A content-based image retrieval system (CBIR) is constructed to provide learner with required information using image recognition and multimedia technologies, such that the objective of m-learning can be achieved. We believe that it is worth developing a mobile learning system to provide the learners with a new educational environment which can recycles the existing PEBs.

Better foreground segmentation for 3D face reconstruction using graph cuts

Research on image-based 3D reconstruction has recently shown a lot of good results, but it assumes precise target objects are already segmented from each input image. Traditionally, background subtraction was used to segment the target objects, but it can yield serious problems, such as noises and holes. To precisely segment the target objects, graph cuts have recently been used. Graph cuts showed good results in many engineering problems, as they can globally minimize energy functions composed of data terms and smooth terms, but it is difficult to automatically obtain prior information necessary for data terms. Depth information generated by stereo vision was used as prior information, which shows good results in their experiments, but it is difficult to calculate depth information for 3D face reconstruction, as the most of faces have homogeneous regions. In this paper, we propose better foreground segmentation method for 3D face reconstruction using graph cuts. The foreground objects are approximately segmented from each background image using background subtraction to assist to estimate data terms of energy functions, and noises and shadows are removed from the segmented objects to reduce errors of prior information. Removing the noises and shadows should cause to lose detail in the foreground silhouette, but smooth terms that assign high costs if neighboring pixels are not similar can fill out the lost silhouette. Consequently, the proposed method can segment more precise target objects by globally minimizing the energy function composed of smooth terms and approximately estimated data terms using graph cuts.

Intelligent document scanning with active camera

Document scanning is important as a prerequisite stage for analysis and recognition. Recently, a lot of researches about document image acquisition using a camera have been attempted, and the camera can be an alternative input device for document scanning if we can solve some problems such as the low resolution. We use an image registration to overcome the low resolution of a camera. An ordinary image registration method needs a pre-processing such as a camera calibration to reduce distortions on the composite. Therefore the ordinary method has an extra running time. In this paper, we propose a component-based image registration method to concentrate on reducing the distortions and acquiring a seamless image using a PTZ (pan-tilt-zoom) camera without pre-processing. Because we divide the input document image into each component and generate the registration on text components using a text-specific characteristic, this method leads to reduce the object (text) distortions on the composite, and we save the extra running time because this method does not perform the post processing.

PDA-based text localization system using client/server architecture

Recently, several research results of image processing are proposed on the mobile vision systems. Many CPUs for Personal Digital Assistant(PDA) are integer CPUs, which have no floating-computation component. It results in slow computation of the algorithms constructed by using neural networks, which have much floating-computation. In this paper, in order to resolve this weakness, we propose an effective text localization system with the Client(PDA)/Server(PC) architecture which is connected to each other with a wireless LAN. The Client(PDA) compresses tentative text localization results in JPEG format for minimizing the transmission time to the Server(PC). The Server(PC) uses both the Multi-Layer Perceptron(MLP)-based texture classifier and Connected Components(CCs)-based filtering for a precise text localization based on the Client(PDA)s tentative extracting results. The proposed method leads to not only faster running time but also efficient text localization.