Sangwook Oh

Relevance graph-based image retrieval

The basic limitation of content-based image retrieval and relevance feedback based on low-level image features is that low-level features are often highly ineffective for representing nor only content similarity, but conceptual and contextual similarity between images. On the other hand, the utility of text-based image retrieval is restricted due to the limited availability of image annotations and textual descriptions limited ability in describing image content. In this paper, we introduce a novel approach to content-, concept- and context-based image retrieval that utilizes user-established relevance between images only using image links without relying on image features or textual annotations. We present a framework for accumulating image relevance information through relevance feedback, determining the degree of relevance, and constructing a relevance graph for an image database. The use of graph-theoretical algorithms is suggested for image search and experimental studies are presented to demonstrate the potential of the proposed methods.

Fast Text Caption Localization on Video Using Visual Rhythm

In this paper, a fast DCT-based algorithm is proposed to efficiently locate text captions embedded on specific areas in a video sequence through visual rhythm, which can be fast constructed by sampling certain portions of a DC image sequence and temporally accumulating the samples along time. Our proposed approach is based on the observations that the text captions carrying important information suitable for indexing often appear on specific areas on video frames, from where sampling strategies are derived for a visual rhythm. Our method then uses a combination of contrast and temporal coherence information on the visual rhythm to detect text frames such that each detected text frame represents consecutive frames containing identical text strings, thus significantly reducing the amount of text frames needed to be examined for text localization from a video sequence. It then utilizes several important properties of text caption to locate the text caption from the detected frames.

Relevance feedback reinforced with semantics accumulation

Relevance feedback (RF) is a mechanism introduced earlier to exploit a user’s perceptual feedback in image retrieval. It refines a query by using the relevance information from the user to improve subsequent retrieval. However, the user’s feedback information is generally lost after a search session terminates. In this paper, we propose an enhanced version of RF, which is designed to accumulate human perceptual responses over time through relevance feedback and to dynamically combine the accumulated high-level relevance information with low-level features to further improve the retrieval effectiveness. Experimental results are presented to demonstrate the potential of the proposed method.

An automated glitch-detection/restoration method of atomic force microscope images

An automated glitch-detection/restoration method of atomic force microscope images is proposed and implemented. Contrary to other manual methods, our method is based on the probability distribution of the derivative of the scanned image data. The glitches are identified as the points that deviate from a normal probability density function. The essence of the automation is calculating the distribution of the scanned image and removing the points that deviate from the normal distribution. Quantitative analysis of the original and the restored image have been performed and the degree of deformation of the restored images has also been analyzed. This technique can directly be applied to other types of scanning probe microscope equipments.

Design of color-code system for time-stamping broadcast video

For the purpose of video indexing and monitoring of broadcast TV streams, it is needed to associate any frame of a broadcast stream with its current broadcasting time. We propose a new time-stamping method to associate any frame of a broadcast stream accurately with its corresponding broadcasting time by stamping a structured color-code visually representing the current broadcasting time on each frame of a broadcast stream. The experimental results show that a structured color-code with brightness calibration can be perfectly recognized in real-time. The robustness of the structured color-code is also demonstrated by implementing a real-time video indexing system that can generate metadata containing segmentation information described by broadcasting time. TV viewers can interactively view broadcast streams recorded in their set-top boxes with the corresponding metadata delivered.

Detection and volume estimation of semiconductor quantum dots from atomic force microscope images

An automated quantum dot detection and volume estimation method of atomic force microscope images are proposed and implemented. Quantum dot detection is based on the image-segmentation method widely used in image processing. We used a modified local watershed algorithm, which produces stable and effective segmentation results in order to detect the peak position of the quantum dot of the atomic force microscopy image. An effective volume estimation of the quantum dot is performed by intelligently determining the boundary of the quantum dot at each height step and by simply adding small volume fractions lying inside of the quantum dot boundary. Our volume estimation results are quantitatively analyzed by comparison with the quantum dot volume obtained by assuming conical and pyramidal shapes. The detection results of our method are quantitatively compared with the results of two other methods for general grain detection, and with a transmission electron micrograph of the quantum dot. Finally, information such as the number of quantum dots, the density, the size and the height distribution are also provided as a result of implementing our method. This method can be applied to detect and estimate volume of a similar shape with a different size.