Michihiro Kobayakawa

Fast Computation of Similarity Based on Jaccard Coefficient for Composition-Based Image Retrieval

This paper proposes an algorithm and data structure for fast computation of similarity based on Jaccard coefficient to retrieve images with regions similar to those of a query image. The similarity measures the degree of overlap between the regions of an image and those of another image. The key idea for fast computation of the similarity is to use the runlength description of an image for computing the number of overlapped pixels between the regions. We present an algorithm and data structure, and do experiments on 30,000 images to evaluate the performance of our algorithm. Experiments showed that the proposed algorithm is 5.49 (2.36) times faster than a naive algorithm on the average (the worst). And we theoretically gave fairly good estimates of the computation time.

A feature independent of bit rate for twinvq audio retrieval

In this paper we propose an audio feature for TwinVQ audio retrieval. For making effective audio database, we consider that these two techniques (compression and feature extraction) are dealt with on one platform. The proposed audio feature satisfies the following requirements: 1) independent of bit rate; 2) extractable from compressed data without decoding; 3) computable in the framework of TwinVQ. We show that the autocorrelation coefficient is theoretically independent of bit rate and confirm experimentally that the feature computed from CD audio data is actually independent of bit rate.

A method for retrieving music data with different bit rates using MPEG-4 TwinVQ audio compression

The present paper describes a method for indexing a piece of music using the TwinVQ (Transform-domain Weighted Interleave Vector Quantization) audio compression (MPEG-4 audio standard). First, we present a framework for indexing a piece of music based on the autocorrelation coefficients computed in the encoding step of TwinVQ audio compression. Second, we propose a new music feature that is robust with respect to bit rate based on the fact that the i-th autocorrelation coefficient with bit rate B1 of a piece of music computed in the encoding step of TwinVQ audio compression can approximate the j-th autocorrelation coefficient with bit rate B2 of the piece of music where i= left lfloor frac B_1 B_2 j right rfloor, and on the wavelet transform. Finally, we perform retrieval experiments on 1,023 pieces of polyphonic music with bit rate (8 kbps, 12 kbps, 16 kbps, 20 kbps, 24 kbps, 28 kbps, 32 kbps, 36 kbps, 40 kbps, and 44 kbps). The experimental results indicate that the proposed music feature for indexing has excellent retrieval performance for queries of various bit rates.

mm-GNAT: index structure for arbitrary L p norm

For fast epsiv-similarity search, various index structures have been proposed. Yi et at. proposed a concept multi-modality support and suggested inequalities by which epsiv-similarity search by L1, L2 and Linfin norm can be realized. We proposed an extended inequality which allows us to realize epsiv-similarity search by arbitrary Lp norm using an index based on Lq norm. In these investigations a search radius of a norm is converted into that of other norm. In this paper, we propose an index structure which allows search by arbitrary Lp norm, called mm-GNAT (multi-modality support GNAT), without extending search radius. The index structure is based on GNAT (geometric near-neighbor access tree). We show that epsiv-similarity search by arbitrary Lp norm is realized on mm-GNAT. In addition, we performed search experiments on mm-GNAT with artificial data and music data. The results show that the search by arbitrary Lp norm is realized and the index structure has good search performance.

A Texture Segmentation Method Using Hierarchical Correlations of Wavelet Coefficients

For making an effective and simple region-based image retrieval system, it needs to uniformly realize both image segmentation and retrieval. In this paper, we focus on texture segmentation for region-based texture retrieval, and propose a new texture segmentation method based on the hierarchical correlation between the wavelet coefficients of adjacent level of wavelet decomposition. Firstly, we define a texture feature which is extracted from the hierarchical relations of wavelet coefficients. Secondly, we propose an algorithm for texture segmentation using the texture feature. Lastly, we evaluate the performance of texture segmentations. Experiments show that our method has a good performance for texture segmentation and suggest that the proposed texture segmentation method is applicable to region-based texture retrieval.

Music Genre Classification of MPEG AAC Audio Data

In this paper, we propose a musical feature extracted from the bit stream of AAC (Advanced Audio Coding) compressed audio data without decoding to audio signals. We focus on the spectral data which are stored in the bit stream for representing the flatten MDCT (Modified Discrete Cosine Transform) of an audio signal. For computing the musical feature, we extract the spectral data and apply the Discrete Wavelet Transform (DWT) to the extracted spectral data. For musical genre classification, we use the discriminant analysis as a classifier. We experimented on 1,498 AAC compressed audio data collected from 10 musical genres and evaluated the performance of the musical feature. We got the maximum correct ratios 81.24%. The experiments showed that the musical feature based on the spectral data in the bit stream had good performance for genre classification in the MPEG-4 AAC compressed domain.

A method for music structure analysis using MPEG-4 TwinVQ audio compression

This present paper propose a method for analyzing a music structure using autocorrelation coefficients computed in the encoding step of TwinVQ audio compression. We phrase the autocorrelation sequence into subsequences by using the extracted musical unit, and then classify the subsequences and assign a label to a class. From a sequence of labels, we extract subsequences of longest match to analyze structure of a piece of music. To evaluate performance of our method, we compare the extracted subsequences by our method with that by hand. The experimental results indicates that our method has a good performance for analyzing structure of a piece of music.

A Partial Retrieval of Music Data with Different Bit Rate using MPEG-4 TwinVQ Audio Compression

The present paper describes a method for partial retrieval of a piece of music using the autocorrelation coefficients computed in the encoding step of TwinVQ (Transform-domain Weighted Interleave Vector Quantization) audio compression (MPEG-4 audio standard). Our key contribution is to realize partial retrieval of a piece of music that is robust with respect to bit rate using an approximation relation. The approximation relation is based on the fact that the i-th autocorrelation coefficient with bit rate B1 of a piece of music computed in the encoding step of TwinVQ audio compression can approximate the j-th autocorrelation coefficient with bit rate B2 of the piece of music, where i = [(B1/B2) j]. First, we present our frame work for music information retrieval and music retrieval of a piece of music based on the autocorrelation coefficients computed in the encoding step of TwinVQ audio compression. Then, we show experimental results of partial retrieval of a piece of music on 1,023 pieces of music. The experimental results indicate that partial retrieval using the autocorrelation coefficients has excellent retrieval performance for queries of various bit rates.