KR20090051430A - Jpeg-2000 compressed image retrieval using partial entropy decoding - Google Patents

Abstract

The present invention relates to a JPEG-2000 compressed image retrieval method using partial entropy decoding. The present invention relates to entropy decoding on an image compressed with JPEG 2000 using statistical characteristics of various contextual information generated through the JPEG-2000 entropy encoding process. By providing a new JPEG-2000 compressed image retrieval method for constructing a feature vector by partially performing the process, the memory and retrieval time for image retrieval are greatly reduced.

Jpeg 2000, search, context, probability, similarity

Description

Jpec-2000 compressed image retrieval using partial entropy decoding

The present invention relates to a JPEG-2000 compressed image retrieval method using partial entropy decoding, and more particularly, to a new JPEG-2000 compressed image retrieval that constructs a feature vector by partially performing an entropy decoding process.

With the rapid development of the Internet, most users have demanded that their service providers deliver high-quality multimedia.

This includes the development of compression technologies such as the Motion Picture Experts Group (MPEG) and Joint Photographic Experts Group (JPEG) / JPEG 2000 and the development of technologies to build a network environment with guaranteed quality of service (QoS). It was an opportunity to actively promote.

MPEG is a standard of compression technology that allows information to be transmitted through video compression and code representation that continuously change over time.

J-Peck is a standard technology for compressing still images such as photographs for communication. It can display 16 million colors and is suitable for high resolution display devices. If you try to compress a large file into a very small file, the quality of the image will be degraded. However, using the JPEG compression technology, you can adjust the image to prevent the image from being damaged.

The basis of this J-Tech technology is a Fourier transform called DCT (Discrete Cosine Transform), which splits an image into blocks of 8 × 8, approximates 64 basic patterns through DCT, and applies coefficient values to these basic patterns. Is quantized by dividing by the largest value, and encoding is performed using Huffman coding, which compresses the quantized value by assigning a smaller value to a frequently appearing bit.

After all, J-PEK's fundamental technology is a method of compressing by lowering the amount of information by unambiguously expressing high frequency components such as hair, which are difficult to distinguish from the human eye in the image. In the opposite direction, the desired image is finally obtained through an inverse discrete cosine transform (IDCT).

In case of JP-2000, bit-plane based coding is performed instead of JP-based pixel-based coding by using DWT instead of DCT and Arithmetic Coding instead of Huffman coding. do.

DWT was developed in discrete form for computers using wavelet transforms. The disadvantage of Fourier transforms, which can represent continuous signals as sums of trigonometric functions, is that the trigonometric functions repeat indefinitely with periods. Because of this, the signal that is hard to respond to the change of time is more distorted than the original shape, but the wavelet transform can express the spatial and frequency components unlike DCT of the existing JEGEC, so that various resolutions can be obtained. Support is available.

In addition, the entire image can be reduced to a lower resolution picture such as 1/4, 1/16, 1/64, etc. When the zoomed out, the high frequency components are stored separately and the original image is restored using the information again when zoomed in. Therefore, the distortion is greatly reduced.

This compressed image is compressed using compression techniques such as run-length coding (RLC) and entropy coding.It is Jpec 2000.It saves the high-frequency components to be used when the original image is reduced to a smaller image and enlarged again. Since these high frequency components are often repeated with the same value, compression can be performed very efficiently by compressing with entropy coding.

As a result, the same image quality can be obtained by compressing 5 times as much as 1/150 compared to 1/30 compression ratio of existing JPEG, which can reduce the overall network load of the Internet and store it on a computer or an important server on the Internet. You can also reduce your storage space by 1/5.

Due to these advantages, a method of describing a feature vector for searching for a texture image in the JPEG-2000 compressed region is required as it is gradually converted from JPEG to JPEG-2000.

However, existing image retrieval algorithms have constructed a feature vector for retrieval in the spatial domain or frequency domain. This requires a lot of time and memory to retrieve compressed images.

The present invention was developed to solve the above problems and the problem of the conventional search method, and entropy decoding in the image compressed with JPEG 2000 using statistical characteristics of various context information generated through the JPEG-2000 entropy encoding process. It is an object of the present invention to significantly reduce the memory and retrieval time for performing image retrieval by providing a new Jpec-2000 compressed image retrieval method constituting a feature vector.

A feature of the present invention for achieving the above object is entropy by selecting a subband for constituting a feature vector of the query image in the decoder with a query image compressed in Jpe2000, and selecting only the stream generated from the selected subband coefficients Decoding process is performed, and context information is extracted by performing the context modeling process of the Jpepec 2000 entropy decoding step, and the occurrence frequency of each context information is examined in the extraction process, and the occurrence frequency of each context information is normalized by probability distribution. Using the probability distribution of the context information generated in each bit plane, the entropy for each bit plane is examined to select a predetermined number of bits in high order, and to select context information for constructing the feature vector of the query image. A query image processing step of extracting a probability value of the context information; The search target image compressed with the JPEG-2000 performs the entropy decoding process using the subband information selected from the query image, extracts the contextual information by performing the context modeling process of the JPEG-2000 entropy decoding step, and extracts the context information. In this process, the frequency of occurrence of each contextual information is examined, and the frequency of occurrence of each contextual information is normalized by a probability distribution, and the search target image extracts probability values of the selected contextual information using the selected bit plane and contextual information from the query image. Processing step; And performing similarity evaluation using the probability distribution of the query image and the database image in which all the search target images are recorded, sorting the images of the database in the order of similarity, and performing the rest of the JPEG-2000 decoding of the image in the predetermined predetermined rank. And a JPEG-2000 compressed image retrieval method using partial entropy decoding comprising a search image output step of outputting a final retrieved image.

In the above configuration, the JPEG-2000 entropy encoding technique analyzes the structural patterns of neighboring neighboring wavelet coefficients and structural patterns of the critical state coefficients and generates a total of 19 kinds of context information. The context information is encoded in the arithmetic encoding process. It provides a model for predicting the probability of a symbol and can describe the local pattern of a texture image well, and can effectively describe the overall characteristics of an image using statistical characteristics such as a histogram. The texture image to be represented can be efficiently defined.

In addition, J-Peck-2000 minimizes the context information (256 kinds) that can be generated to simplify the arithmetic encoder and optimize the model for each coding technique. According to the structural patterns of the surrounding critical states, 9 of the 256 types of ZC and MC technologies, and 5 of 16 of the SC technologies are summarized in context information. Therefore, from the point of view of image retrieval, the effect can be less affected by the significant state change of the adjacent coefficient due to distortion. Therefore, since the context information may not be very sensitive to the critical state change of the adjacent coefficient bits, it may exhibit robust performance in searching a distorted image database such as various signal processing and geometric changes. In addition, since the probability information of the context information does not change significantly even in the search result using various distortion images as the query image, the flexibility of the present invention can be confirmed, and the performance of the similar image search having a certain pattern is excellent.

In addition, since the present invention partially performs the decoding process, it is analyzed to be superior to the existing invention in search time and memory requirements for constructing the feature vector. This means that the larger the database size, the more the gain of the proposed algorithm increases.

According to the present invention having the configuration as described above, the time and memory is greatly reduced in the search of the image in the environment using the Jpeek 2000 compression method, and also has an excellent search effect in the search for similar and distorted images .

Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings.

FIG. 1 is a block diagram showing a system configuration for implementing a method of searching a JPEG-2000 compressed image. A user converts the JPEG-2000 compressed image into a query image using an input means such as a keyboard or a mouse of a computer. When inputted, the computer searches the JPEG compressed image database and outputs the result through the monitor. The method will be described in detail below.

The compressed video of the JPEG-2000 is first rearranged in a desired form by the bit-stream organization of the decoder, which is a peculiar characteristic of the JPEG-2000. Assuming that the encoding is performed in the compression form of (Quality Scalability), in the JPEG-2000 decoding step, the bit stream may be rearranged to perform resolution scalability in various resolution forms.

In addition, a subband for selecting a feature vector of a query image is selected by a band selection for feature extraction, and only a stream generated from the selected subband coefficient is selected to perform an entropy decoding process.

Context histogram construction is performed by extracting the context information from the partial inverse arithmetic code (Tier-1 Decoding) by performing the context modeling process of the JPEG-2000 entropy decoding step. Examine the frequency of occurrence of each contextual information.

After that, the probability distribution normalization normalizes the frequency of occurrence of each context information with a probability distribution, and the context entropy measurement for bit-plane selection uses the probability distribution of the context information generated in each bit plane. By examining entropy for each bit plane, several bits are selected according to a predetermined number in order of high entropy.

Context selection for feature extraction selects the context information for constructing the feature vector of the query image, and extracts the probability value of the context information selected by the context probability extraction.

In addition, a band selection for feature extraction that processes the query image after rearranging the database image compressed with the JPEG-2000 in a desired form by the bit-stream organization of the decoder. An entropy decoding process is performed using the selected subband information.

In addition, the context information extraction unit (Partial inverse arithmetic code (Tier-1 Decoding)) to perform the context modeling process of the Jpec-2000 entropy decoding step to extract and extract the context information, the context histogram construction ), The frequency of occurrence of each contextual information is examined.

After that, the probability distribution normalization normalizes the frequency of occurrence of each context information with a probability distribution, and uses the bit plane and context information selected in the query image to determine the probability value of the context information selected by the probability probability extraction unit. Extract

Similarity evaluation is performed in a similarity measurement unit using a probability distribution of a query image and all database images, and then all database images are sorted in a similarity order in a database image ranking ordering.

After that, images outside the predetermined rank are discarded by the DB Image In Rank R ?, and the remaining arithmetic decoding and decompression are performed by the residual arithmetic decoding and decompression. Then, the final retrieved image is output.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a block diagram showing a system configuration for implementing a JPEG-2000 compressed image search method.

Claims (1)

In the image search method of JP-2000, Selecting a subband for constituting a feature vector of the query image from the decoder in the MPEG 2000 compressed image, and performing an entropy decoding process by selecting only the stream generated from the selected subband coefficients, the JPEG 2000 entropy decoding step Contextual modeling is performed to extract contextual information, and the frequency of occurrence of each contextual information is investigated in the extraction process, the frequency of occurrence of each contextual information is normalized by probability distribution, and the probability distribution of contextual information generated in each bitplane. Query image processing step of selecting entropy for each bit plane by selecting a predetermined number of bits in high order, selecting context information for configuring a feature vector of the query image, and extracting probability values of the selected context information. Wow; The search target image compressed with the JPEG-2000 performs an entropy decoding process using subband information selected from the query image, and extracts and extracts context information by performing a context modeling process of the JPEG-2000 entropy decoding step. Search process image processing to investigate the frequency of occurrence of each contextual information, normalize the frequency of occurrence of each contextual information by probability distribution, and extract the probability value of the selected contextual information using the selected bitplane and contextual information from the query image. step; And Similarity evaluation is performed using the probability distribution of the query image and the database image in which all the search target images are recorded, and the images of the database are sorted in the order of similarity, and the remaining JP-2000 decoding process is performed on the image in the predetermined predetermined rank. A search image output step of outputting the final searched image; J-PEC-2000 compressed image searching method using partial entropy decoding, characterized in that consisting of.

Images