Hyung-Jeong Yang

Automatic image captioning

We examine the problem of automatic image captioning. Given a training set of captioned images, we want to discover correlations between image features and keywords, so that we can automatically find good keywords for a new image. We experiment thoroughly with multiple design alternatives on large datasets of various content styles, and our proposed methods achieve up to a 45% relative improvement on captioning accuracy over the state of the art.

Automatic detection and recognition of Korean text in outdoor signboard images

In this paper, an automatic translation system for Korean signboard images is described. The system includes detection and extraction of text for the recognition and translation of shop names into English. It deals with impediments caused by different font styles and font sizes, as well as illumination changes and noise effects. Firstly, the text region is extracted by an edge-histogram, and the text is binarized by clustering. Secondly, the extracted text is divided into individual characters, which are recognized by using a minimum distance classifier. A shape-based statistical feature is adopted, which is adequate for Korean character recognition, and candidates of the recognition results are generated for each character. The final translation step incorporates the database of shop names, to obtain the most probable result from the list of candidates. The system has been implemented in a mobile phone and is demonstrated to show acceptable performance.

Coercively Adjusted Auto Regression Model for Forecasting in Epilepsy EEG

Recently, data with complex characteristics such as epilepsy electroencephalography (EEG) time series has emerged. Epilepsy EEG data has special characteristics including nonlinearity, nonnormality, and nonperiodicity. Therefore, it is important to find a suitable forecasting method that covers these special characteristics. In this paper, we propose a coercively adjusted autoregression (CA-AR) method that forecasts future values from a multivariable epilepsy EEG time series. We use the technique of random coefficients, which forcefully adjusts the coefficients with −1 and 1. The fractal dimension is used to determine the order of the CA-AR model. We applied the CA-AR method reflecting special characteristics of data to forecast the future value of epilepsy EEG data. Experimental results show that when compared to previous methods, the proposed method can forecast faster and accurately.

MMSS: multi-modal story-oriented video summarization

We propose multi-modal story-oriented video summarization (MMSS) which, unlike previous works that use fine-tuned, domain-specific heuristics, provides a domain-independent, graph-based framework. MMSS uncovers correlation between information of different modalities which gives meaningful story-oriented news video summaries. MMSS can also be applied for video retrieval, giving performance that matches the best traditional retrieval techniques (OKAPI and LSI), with no fine-tuned heuristics such as tf/idf.

Hidden pattern discovery on event related potential EEG signals

EEG signals are important to capture brain disorders. They are useful for analyzing the cognitive activity of the brain and diagnosing types of seizure and potential mental health problems. The Event Related Potential can be measured through the EEG signal. However, it is always difficult to interpret due to its low amplitude and sensitivity to changes of the mental activity. In this paper, we propose a novel approach to incrementally detect the pattern of this kind of EEG signal. This approach successfully summarizes the whole stream of the EEG signal by finding the correlations across the electrodes and discriminates the signals corresponding to various tasks into different patterns. It is also able to detect the transition period between different EEG signals and identify the electrodes which contribute the most to these signals. The experimental results show that the proposed method allows the significant meaning of the EEG signal to be obtained from the extracted pattern.

Assembly design ontology for service-oriented design collaboration

AbstractThis paper presents an Assembly Design (AsD) ontology that explicitly represents AsD constraints and infers any remaining implicit ones. By relating concepts through ontology technology rather than just defining data syntax, assembly and joining concepts can be captured in their entirety or extended as necessary. Ontologies allow assembly and joining constraints to be represented in a standard manner regardless of geometry file formats. Such representation will significantly improve service-oriented design collaboration. The developed AsD ontology is tested using a realistic mechanical assembly and it is shown how the ontology can be used to capture design rationale and analyze the design intents. In conclusion, the significance of ontology for realizing lean and selective assembly design information sharing is discussed.

Automatic Detection and Recognition of Shop Name in Outdoor Signboard Images

In this paper, a system for automatic detection and recognition of Korean texts or shop names in outdoor signboard images is described. The system includes detection, binarization and extraction of text in a signboard image captured by a camera of a mobile phone for the recognition of the shop name. It can deal with different font styles and sizes as well as illumination changes. Individual characters detected by connected component analysis are recognized by using nonlinear mesh, in which feature vectors of vertical and horizontal components are extracted from the binarized image. Proposed methods have been applied to a Korean text translation system, which can automatically detect and recognize Korean texts and generate the translation result.

Multimodal learning using convolution neural network and Sparse Autoencoder

In the last decade, pattern recognition methods using neuroimaging data for the diagnosis of Alzheimers disease (AD) have been the subject of extensive research. Deep learning has recently been a great interest in AD classification. Previous works had done almost on single modality dataset, such as Magnetic Resonance Imaging (MRI) or Positron Emission Tomography (PET), shown high performances. However, identifying the distinctions between Alzheimers brain data and healthy brain data in older adults (age > 75) is challenging due to highly similar brain patterns and image intensities. The corporation of multimodalities can solve this issue since it discovers and uses the further complementary of hidden biomarkers from other modalities instead of only one, which itself cannot provide. We therefore propose a deep learning method on fusion multimodalities. In details, our approach includes Sparse Autoencoder (SAE) and convolution neural network (CNN) train and test on combined PET-MRI data to diagnose the disease status of a patient. We focus on advantages of multimodalities to help providing complementary information than only one, lead to improve classification accuracy. We conducted experiments in a dataset of 1272 scans from ADNI study, the proposed method can achieve a classification accuracy of 90% between AD patients and healthy controls, demonstrate the improvement than using only one modality.

A novel and effective method for specular detection and removal by tensor voting

Most specular detection methods assumed that dominant highlight regions should be uniform for the detection of highlights, which may not be the case in real images. Even when non-uniformity is allowed in the detection, the specular removal can still suffer from non-converged artifacts due to discontinuities in surface colors, especially in highly textured and multicolor images. In this paper, we propose a novel and effective resolution to separate and remove specular components from a single image by adopting tensor voting to obtain reflectance distribution of an input image. Specular and noise pixels denoted as small tensors are isolated and removed. Diffuse reflectance distribution is achieved by analyzing salient and orientation information of tensors around the specular region. The proposed method is non-iterative and does not require any predefined constraints in the input image. We evaluate our proposed method on a dataset consisting of highly textured and multicolor images. Experimental results showed that our result is outstanding compared to other state-of-the-art techniques.

Recognition of Text in Wine Label Images

In this paper, an automatic recognition system for Wine label images is described. The system includes detection and extraction of text for the recognition for Wine label images. It deals with impediments caused by different font styles and font sizes, as well as illumination changes and noise effects. Firstly, the text region is extracted by an edge-histogram, and the text is binarized by clustering. Secondly, the extracted text is divided into individual characters, which are recognized by using the Multi-Layer Perceptron. A shape-based statistical feature is adopted and the recognition results are generated for each character. The system has been implemented in a mobile phone and is demonstrated to show an acceptable performance.