An improved Bag-of-Words framework for remote sensing image retrieval in large-scale image databases

Figures & data

Figure 1. Landsat-5 TM archive coverage.

Figure 2. Architecture of RSIR framework.

Figure 3. Distance ratio: most incorrect matches have values of this ratio that is greater than 0.9.

Figure 4. Visual word assignment process of our approach. (a) Target image, (b) base image (reference image).

Figure 5. The RS image representation method based on features from the base image. First the image match process between the target image and base image was performed then the corrected match point was recorded with a sequence array.

Figure 6. The similarity between a pairs of images is determined by the size of their overlapping regions in the visual word feature space of the base images. A: region of a query image in feature space; B: region of a target image in feature space.

Figure 7. Indirect comparison of image 1 and image 2 with a template image as a common comparator.

Figure 8. Storage requirements: the BoW framework vs. our approach.

Figure 9. Some examples of RS image in the data-set used in the experiment.

Figure 10. Some examples of query images.

Table 1. Comparison of occupied storage space for FR, BoW, and our approach.

Database	Image number	Descriptors	Number of visual word	Occupied storage space (byte)
FR	8223	12.3 M		1.5 G
BoW			10 K	34 M
			100 K	45 M
			200 K	58 M
Ours (1BI)			22 K	5 M
Ours (2BI)			45 K	8 M
Ours (3BI)			66 K	12 M
Ours (4BI)			92 K	16 M

Table 2. Comparison of the performance of the Top 20 scores for BoW framework and the RSIR framework.

	Query image
Query method	Landsat-7	HJ-1A	SPOT4
BoW (10K)	0.36	0.31	0.32
BoW (100K)	0.43	0.37	0.34
BoW (200K)	0.44	0.38	0.35
Ours	1	1	1

Table 3. Comparison between the AP scores of the BoW framework and the approach used in this study.

	Query image
	Landsat-7		HJ-1A		SPOT4
Query method	Hard	Soft	Hard	Soft	Hard	Soft
BoW (10K)	0.36	0.37	0.36	0.36	0.33	0.35
BoW (100K)	0.39	0.41	0.37	0.39	0.36	0.38
BoW (200K)	0.41	0.42	0.37	0.39	0.36	0.39
Ours (1BI)	0.60		0.51		0.53
Ours (2BI)	0.64		0.61		0.62
Ours (3BI)	0.67		0.62		0.65
Ours (4BI)	0.67		0.63		0.65

Figure 11. Comparison of query results for the BoW framework and the approach used in this study. The WRS path/row of the query image was 121/39. The image in the second line gives the results returned using our approach. The query image is the ROI within the blue rectangle in the image. All the returned images for our approach were positive answers. The third line shows the results retrieved using the BoW framework. Rank 3 and rank 5 in this line are negative answers.

Figure 12. Precision recall graph: comparison between the BoW framework and our approach.