Unified feature analysis in JPEG and JPEG 2000-compressed domains

Abstract

Retrieving images compressed by different algorithms typically involves a pre-processing operation to decompress them onto the spatial domain from which features are extracted for further analysis. Our objective is to investigate common features that can be found in JPEG-compressed and JPEG 2000-compressed images so that image indexing can be done directly in their respective compressed domains. A fundamental difference between JPEG and JPEG 2000 is their transforms; the former uses a block-based discrete cosine transform (BDCT) while the latter uses a wavelet transform (WT). Direct comparison on BDCT blocks and WT subbands cannot reveal their relationship. By employing our proposed subband-filtering model, the BDCT coefficients can be concatenated to form structures similar to WT subbands. Our theoretical studies show that the concatenated BDCT and WT filters share common characteristics in terms of passband regions, magnitude and energy spectra. In particular, their low-pass filters are identical for Haar wavelets and highly similar for other wavelet kernels. Despite the fact that compression can affect features that can be extracted, our experimental results confirm that common features can always be extracted from JPEG- and JPEG 2000-compressed domains irrespective of the values of the compression ratio and the types of WT kernels used. As a result, similar JPEG-compressed and JPEG 2000-compressed images can be retrieved from one another without requiring a full decompression.

Introduction

With the rapid growth of Internet and multimedia systems, the use of visual information has increased enormously, such that image indexing and retrieval techniques have become important. These techniques extract meaningful information (features) from an image so that images can be classified and retrieved efficiently based on their contents. Examples of application areas include video on demand applications and multimedia information systems.

Over recent decades, retrieval systems such as QBIC, Virage, Photobook and Visual SEEK have been proposed [1], [2], [3], [4], [5], [6], [7], [8]. All of these systems use features extracted from the spatial domain for indexing. This is in contrast to the fact that images are often compressed using JPEG or JPEG 2000 to reduce their size for storage and transmission [9], [10], [11]. Retrieving these kinds of compressed images then requires reconversion to the uncompressed spatial domain for feature extraction and analysis. This approach requires many decompression operations, especially for large image archives. To avoid some of these operations, it was proposed that feature extraction be done directly in the transformed domains. For example, DCT-based indexing techniques extracted DCT coefficients in JPEG for indexing [12], [13], [14], [15], [16], [17], [18], [19] while wavelet-based indexing techniques used wavelet coefficients as features [20], [21], [22], [23], [24], [25], [26].

Different compression techniques result in different transformed coefficients. This in turn implies that features that can be extracted in the compressed domain depend greatly on the compression scheme used. It is extremely desirable if we can explore features that are common in the uncompressed spatial domain, DCT and wavelet transform (WT) domains, so that indexing can be done from their original domain without incurring a full decompression. Direct comparison between DCT and WT domains cannot easily reveal their relationship. We proposed a subband-filtering model that can be used to theoretically study transformed outputs from JPEG and JPEG 2000 schemes. This helps quantify how similar the two transformed outputs are and thus helps identify common features in these two transformed domains. As compression can seriously affect image quality, it will also affect features that can be extracted in the compressed domains. To test the effect of compression on the common features found in the two transformed domains, a retrieval system was built that consisted of JPEG-compressed and JPEG 2000-compressed images at different compression ratios.

This paper is organized as follows. Section 2 briefly discusses the DCT in JPEG and the WT in JPEG 2000 compression schemes. Our proposed subband-filtering model used to relate DCT and WT is described in Section 3. Using this model, similarity between DCT and WT subbands can be discovered and is discussed in Section 4. Section 5 then studies the effect of compression on the subbands similarity, while Section 6 describes features that are extracted in different compressed domains for indexing. Our experimental results on retrieving JPEG-compressed and JPEG 2000-compressed images are given in Section 7. Finally, Section 8 concludes the paper.