Unified feature analysis in multiple compressed domains

Abstract

Visual information often requires huge storage space. Owing to the constraints of transmission and storage, images are usually compressed by either JPEG or JPEG2000 to reduce its storage requirement. Over the past decade, retrieval systems operated in uncompressed/spatial and compressed domains have been proposed. However, they all function in single domain only. Retrieving these kinds of images in multiple domains typically requires reconverting them into spatial domain in which features are extracted for further analysis. This approach incurs many pre-processing operations such as decompression, especially for large image archives. The objective of this study is to investigate the common features in different compressed domains so that image indexing can be done directly from their respective domains. A fundamental difference between JPEG and JPEG2000 is their transformation schemes. JPEG and JPEG2000 employ dissimilar Block-based Discrete Cosine Transform (BDCT) and Wavelet Transform (WT) respectively. Direct comparison on BDCT blocks and WT subbands cannot expose their relationship. By employing a subband filtering model, filters in BDCT and WT can be directly compared. In accordance with our intensive mathematical analysis, BDCT coefficients can be concatenated to form structures similar to WT subband. Under the same structure, BDCT and WT filters are comparable. Considering JPEG2000 Part I and II compression schemes, commonly used wavelet kernels are involved in our comparison. Theoretical studies show that both BDCT and WT filters share common characteristics in terms of passband region, magnitude and energy spectra. Particularly, lowpass filters in the two transforms are the same for Haar wavelet. In addition, both lowpass and bandpass filters of the selected kernels provide high similarities. Outputs of the two subband models are alike. Common features can thus be extracted from the BDCT and WT subband outputs. Though high similarities are confirmed between BDCT and WT outputs, compression may affect their similarities. This is because the compression schemes of JPEG and JPEG2000 are quite different starting from transformation to quantization. The effect of compression on their similarities is worth examining. A variety of images are compressed at various compression ratios ranging from 1.6:1 to 72:1 by BDCT in JPEG and different wavelet kernels in JPEG2000. Studies on their output spectra expose their similarities under compression. Despite of high compression, large similarities can still be found. To validate our proposed subband filtering model, an image retrieval system is established. The system aims to search for images in different compressed domains by applying the model to partially decompressed images. To overcome the effect of shifting, scaling and rotation, translation and rotation invariant features are extracted from the images. Our simulation results present high precision and recall values at all compression ratios. Our proposed indexing algorithm concludes that relevant images can be searched from different compressed domains, regardless of the wavelet kernel or compression ratio used. Both theoretical and experimental studies confirm that common features can be extracted directly from multiple compressed domains irrespective of the value of the compression ratio and the use of BDCT and WT kernels. Relevant JPEG and JPEG2000 images can be retrieved from one and the other without incurring full decompression.