Yunju Baek

Re-ranking algorithm using post-retrieval clustering for content-based image retrieval

In this paper, we propose a re-ranking algorithm using post-retrieval clustering for content-based image retrieval (CBIR). In conventional CBIR systems, it is often observed that images visually dissimilar to a query image are ranked high in retrieval results. To remedy this problem, we utilize the similarity relationship of the retrieved results via post-retrieval clustering. In the first step of our method, images are retrieved using visual features such as color histogram. Next, the retrieved images are analyzed using hierarchical agglomerative clustering methods (HACM) and the rank of the results is adjusted according to the distance of a cluster from a query. In addition, we analyze the effects of clustering methods, querycluster similarity functions, and weighting factors in the proposed method. We conducted a number of experiments using several clustering methods and cluster parameters. Experimental results show that the proposed method achieves an improvement of retrieval effectiveness of over 10% on average in the average normalized modified retrieval rank (ANMRR) measure.

An efficient block-matching criterion for motion estimation and its VLSI implementation

The block-matching motion estimation is the most popular technique for motion compensated coding of image sequence. Due to the intensive computational requirement to perform motion estimation (ME) in real-time, application specific VLSI implementation of the ME is indispensable. We present a novel block-matching criterion for motion estimation called reduced bits mean absolute difference (RBMAD). By comparison with conventional schemes, our scheme reduces hardware requirement and increases the speed of computation in VLSI chip with acceptable video performance. We describe in detail the video performances of proposed criterion and conventional ones. We also show our VLSI implementation using the proposed scheme to compare the hardware requirement and operating speed with conventional ones. It is found that RBMAD using 4 bits has reasonable video performance with 57% less VLSI area and 34% faster, thus it is suitable for low cost applications of video coding.

Majority based ranking approach in web image retrieval

In this paper, we address a ranking problem in web image retrieval. Due to the growing availability of web images, comprehensive retrieval of web images has been expected. Conventional systems for web image retrieval are based on keyword- based retrieval. However, we often find undesirable retrieval results from the keyword based web image retrieval system since the system uses the limited and inaccurate text information of web images ; a typical system uses text information such as surrounding texts and/or image filenames, etc. To alleviate this situation, we propose a new ranking approach which is the integration of results of text and image content via analyzing the retrieved results. We define four ranking methods based on the image contents analysis of the retrieved images; (1) majority-first method, (2) centroid-of-all method, (3) centroid-of-top K method, and (4) centroid-of-largest-cluster method. We evaluate the retrieval performance of our methods and conventional one using precision and recall graphs. The experimental results show that the proposed methods are more effective than conventional keyword-based retrieval methods.

A Ranking Algorithm Using Dynamic Clustering for Content-Based Image Retrieval

In this paper, we propose a ranking algorithm using dynamic clustering for content-based image retrieval(CBIR). In conventional CBIR systems, it is often observed that visually dissimilar images to the query image are located at high ranking. To remedy this problem, we utilize similarity relationship of retrieved results via dynamic clustering. In the first step of our method, images are retrieved using visual feature such as color histogram, etc. Next, the retrieved images are analyzed using a HACM(Hierarchical Agglomerative Clustering Method) and the ranking of results is adjusted according to distance from a cluster representative to a query.We show the experimental results based on MPEG-7 color test images. According to our experiments, the proposed method achieves more than 10 % improvements of retrieval effectiveness in ANMRR(Average Normalized Modified Retrieval Rank) performance measure.

Rotation-, scale-, and translation-invariant image watermark using higher order spectra

Digital watermarks offer means of protecting the copyright of digital multimedia. However, many of the proposed watermarking meth- ods are vulnerable to the geometrical distortions that occur during nor- mal use of the media. We propose a new image watermarking method that is resilient to rotation, scaling, and translation (RST). We use the higher order spectra (HOS), in particular, the bispectrum feature vector of an image as the watermark. The bispectrum is the Fourier spectrum of the triple correlation of a signal. Phases of the integrated bispectra are invariant to translation and scaling. Rotation invariance is achieved using the Radon transform of the image. An image is decomposed into the 1-D projections and we construct a feature vector from them. A watermark is embedded by modifying the vector. We measure the distance between the feature vector extracted from the test image and the watermark at detector. Results of experimental studies show that our method is robust to geometric attacks, JPEG compression, and Gaussian noise.

Web image retrieval using majority-based ranking approach

Web image retrieval has characteristics different from typical content-based image retrieval; web images have associated textual cues. However, a web image retrieval system often yields undesirable results, because it uses limited text information such as surrounding text, URLs, and image filenames. In this paper, we propose a new approach to retrieval, which uses the image content of retrieved results without relying on assistance from the user. Our basic hypothesis is that more popular images have a higher probability of being the ones that the user wishes to retrieve. According to this hypothesis, we propose a retrieval approach that is based on a majority of the images under consideration. We define four methods for finding the visual features of majority of images; (1) majority-first method, (2) centroid-of-all method, (3) centroid-of-top K method, and (4) centroid-of-largest-cluster method. In addition, we implement a graph/picture classifier for improving the effectiveness of web image retrieval. We evaluate the retrieval effectiveness of both our methods and conventional ones by using precision and recall graphs. Experimental results show that the proposed methods are more effective than conventional keyword-based retrieval methods.

Robust Image Watermark Using Radon Transform and Bispectrum Invariants

Image watermark that is resistant to geometric distortion is remained to be an unsolved problem. Difficulty of the probelm comes from the situation that the watermark should be extracted without any information of the original image. In this paper, we review this problem and propose a new watermarking scheme based on invariant pattern recognition theory. We propose an invariant watermark using the Radon transform and higher order spectra. A bispectrum feature vector of the image is used as the watermark. Our approach differs from the previous methods in that we embed watermark into the phase of the higher order spectra. Also, our Radon embedding grid outperforms the Fourier-Mellin based methods. We devised a new embedding method which allows detection of the watermark when there is no exact inverse function during embedding. As we use the Radon transform, our method can be used for medical images. We show the invariance of the designed watermark with mathematical proofs. Experimental results confirm that this scheme is resistant to geometric distortions.

Adaptive digital image watermarking using variable size of blocks in frequency domain

An adaptive digital image watermarking system considering of the human visual system (HVS) is presented in the frequency domain. The digital watermarks are embedded in selected discrete cosine transform (DCT) coefficients of variable size of blocks. To accommodate the HVS characteristics, the image is divided into variable size of blocks, where each block contains only homogeneous regions. The watermark is inserted by significantly changing the selected DCT coefficients of the block that has great variance or by slightly changing for the stationary block. It is shown that the proposed method has good robustness against several image processing operations without significant degradation of image quality by simulation using various test images.

Adaptive bit-reduced mean absolute difference criterion for block-matching algorithm and its VLSI design

An adaptive bit-reduced mean absolute difference (ABRMAD) is presented as a criterion for the block-matching algorithm (BMA) to reduce the complexity of the very large scale integration (VLSI) imple- mentation and to improve the processing time. The ABRMAD uses the lower pixel resolution of the significant bits instead of full-resolution pixel values to estimate the motion vector (MV) by examining the pixels in a block. Simulation results show that the 4-bit ABRMAD has competitive mean square error (MSE) results and a half less hardware complexity than the mean absolute difference (MAD) criterion. It has also better characteristics in terms of both MSE performance and hardware com- plexity than the MiniMax criterion and has better MSE performance than difference pixel counting (DPC), binary block matching with edge map (BBME), and bit-plane matching (BPM) with the same number of bits.

A robust watermarking method for high quality image

We propose a new image watermarking method for copyright protection. This method is a block based approach in the spatial domain. In addition to the perceptual invisibility and robustness, the image quality is also considered during the embedding procedure. Block activity is calculated to embed watermark. The watermark strength and the spread of the signal into a block are controlled by the block activity measurement. Experimental results show that the proposed method is suitable for the high quality image.