JPEG Compressed Image Retrieval via Statistical Features Source: Pattern Recognition, Vol. 36, No. 4, April 2003, pp. 977-985 Authors: Guocan Feng, Jianmin Jiang Speaker: Wen-Chuan Wu Date: 2004/04/19
Outline Motivation Proposed scheme Experiments Conclusions Key construction Similarity measurement Experiments Conclusions
Motivation Compressed image retrieval Query JPEG Images Decompress, Feature extract, similarity Time consuming, computationally expensive JPEG Images
Motivation (conts) DCT domain low frequency DCT IDCT
Proposed scheme Moment extraction DCT: IDCT:
Proposed scheme (conts) Example:
Proposed scheme (conts) Compressed image retrieval DCT coefficients
Proposed scheme (conts) Key construction 0~17 18~ 35 36~53 ……. ….…. 93~111 112~127 (σ) 0~63 64~127 128~191 192~255 (m) 6 5 4 3 1 2 (m,σ)
Proposed scheme (conts) Similarity Measurement Database Query WDT
Experiments Evaluate the performance of a CBIR system: Retrieval efficiency database:500 JPEG compressed image Takes 23 sec. without IDCT Takes 51 sec. with IDCT Retrieval effectiveness Test invariance to the transforms Test average rank of all relevant images Test which rank is the first relevant image
Experiments (conts) Retrieval efficiency
Experiments (conts) Retrieval effectiveness - test invariance to the transforms Database: 100 JPEG compressed images Query
Experiments (conts) Retrieval effectiveness - average rank of all relevant images Database: 100 images AVRR : the average rank of all relevant images. IAVRR :the ideal average rank. ex: IAVRR=(8-1)/2=3.5 Ratio of AVRR = AVRR / IAVRR. ex: ratio=(39/8)/3.5=1.39 The best: Ratio of AVRR =1
Experiments (conts) Retrieval effectiveness - which rank is first relevant image Database: 524 JPEG compressed images Query images: 32
Conclusions Retrieval by directly computing the first and second moments from DCT domain. Saving computing cost and storage spaces. Experiments show the retrieval efficiency and effectiveness of the proposed scheme.