Content based image retrieval using MPEG-7 dominant color descriptor

使用 MPEG-7 主色描述子的內容圖像檢索

Student thesis: Master's Thesis

View graph of relations


  • Ka Man WONG

Related Research Unit(s)


Awarding Institution
Award date4 Oct 2004


Content Based Image Retrieval (CBIR) is an automatic process for searching relevant images based on image features and user inputs. MPEG-7 is an international standard for multimedia content description, and it is an important achievement for CBIR. MPEG-7 has a collection of effective descriptors for images, videos, audios and other multimedia contents. In its visual part, several color descriptors are defined, in which Dominant Color Descriptor (DCD) is a compact and effective descriptor. However, due to the shortcomings of its similarity measure method, the DCD’s performance is worse than a more compact Color Layout Descriptor. This is mainly due to the use of a modified Quadratic Histogram Distance Measure (DCD-QHDM) in DCD. In this thesis, it will be shown that the DCD-QHDM is unable to control the upper bound of the distance between two descriptors and its similarity coefficients are unable to utilize color distance to fine-tune the distance. On the other hand, DCD cannot be directly used in Relevance Feedback (RF) image retrieval with simple histogram weighting technique for further improving the accuracy using user feedback. To address these problems, a Merged Palette Histogram (MPH) approach for DCD is proposed in this thesis. For the first problem, a new Merged Palette Histogram Similarity Measure (MPHSM) is proposed to provide an alternative to the DCD-QHDM. This new similarity measure is based on a merged color palette as a common color space to define histogram intersection. Experimental results show that the proposed MPHSM provides better DCD-based retrieval accuracy in terms of the Averaged Normalized Modified Retrieval Rate (ANMRR) and better visual results using different image datasets. To further improve the retrieval accuracy, MPH based Relevance Feedback (MPH-RF) technique is developed for the second problem. The major problem of the conventional histogram weighting RF method is that it is required to apply on a common feature space among all selected relevant images. However, this is not practical for DCD as it uses individual color space. The proposed MPH-RF solving this problem by merging colors and their percentages to form a new DCD, which represents the images in a compact and natural way. Experimental results show that the improvement of MPH-RF is about 0.0568-0.0828 in terms of ANMRR for different combinations of similarity measures and datasets. These results also demonstrated that the combination of MPHSM and MPH-RF techniques making the DCD as accurate as other non-compact MPEG-7 color descriptors such as Scalable Color Descriptor. To demonstrate the practical applications of the proposed algorithms, a web-based CBIR system called MPEG-7 Image Retrieval Refinement based On Relevance feedback (MIRROR) is developed on a Linux platform. The system is based on a set of MPEG-7 standard descriptors and RF functions. MIRROR is designed as a platform for experiments and further developments on MPEG-7 researches. A comprehensive evaluation for comparing performance between different MPEG-7 visual descriptors using the developed system is also reported in this thesis.

    Research areas

  • Optical storage devices, Image processing, MPEG (Video coding standard), Digital techniques, Database management