一個在行動通訊環境下基於MPEG-7與JPEG2000標準之內容擷取系統

Abstract

We present an image retrieval system (IRSystem) with scalable image compression and content-based retrieval for mobile applications. To fit into the mobile channels, JPEG2000 is used to compress the images o reduce the transmission bandwidth and provide scalable image quality. The content-based search takes statistical features of images for syntactical similarity matching and communicates with users to facilitate semantic similarity matching. In the IRSystem, the syntactical features are translated into MPEG-7 descriptors for speedy image retrieval. For the retrieval of natural images, we use two MPEG-7 descriptors including Scalable Color Descriptor (SCD) and Homogeneous Texture Descriptor (HTD) to search for the similar color and texture images in the server database. In our system, the SCD and HTD descriptors are integrated in three proposed approaches, namely, Average Rank Saturated (AvgRankSatur), Dictionary-Sorted (DictSort), and Minimal Rank (MinRank). If the retrieval results are not satisfactory, the system can be further improved using relevance feedback to refine the query from users. We propose a query refinement method, namely, Integrated Minimal Distance Rank (IntegMDR) to increase the overall retrieval rate upon refinement. Our experimental results show that the combined search schemes can achieve the retrieval accuracy in ANMRR of value 0.1 at a very low bit rate of 0.4 bits per pixel. In addition, the query refinement method achieves high retrieval rate in 2 to 3 times queries.

We present an image retrieval system (IRSystem) with scalable image compression and content-based retrieval for mobile applications. To fit into the mobile channels, JPEG2000 is used to compress the images o reduce the transmission bandwidth and provide scalable image quality. The content-based search takes statistical features of images for syntactical similarity matching and communicates with users to facilitate semantic similarity matching. In the IRSystem, the syntactical features are translated into MPEG-7 descriptors for speedy image retrieval. For the retrieval of natural images, we use two MPEG-7 descriptors including Scalable Color Descriptor (SCD) and Homogeneous Texture Descriptor (HTD) to search for the similar color and texture images in the server database. In our system, the SCD and HTD descriptors are integrated in three proposed approaches, namely, Average Rank Saturated (AvgRankSatur), Dictionary-Sorted (DictSort), and Minimal Rank (MinRank). If the retrieval results are not satisfactory, the system can be further improved using relevance feedback to refine the query from users. We propose a query refinement method, namely, Integrated Minimal Distance Rank (IntegMDR) to increase the overall retrieval rate upon refinement. Our experimental results show that the combined search schemes can achieve the retrieval accuracy in ANMRR of value 0.1 at a very low bit rate of 0.4 bits per pixel. In addition, the query refinement method achieves high retrieval rate in 2 to 3 times queries.