Limits...
A Multispectral Image Creating Method for a New Airborne Four-Camera System with Different Bandpass Filters.

Li H, Zhang A, Hu S - Sensors (Basel) (2015)

Bottom Line: For this multispectral system, an automatic multispectral data composing method was proposed.For the difficult registration problem between visible band images and near-infrared band images in cases lacking manmade objects, we presented an effective method based on the structural characteristics of the system.Experiments show that our method can acquire high quality multispectral images and the band-to-band alignment error of the composed multiple spectral images is less than 2.5 pixels.

View Article: PubMed Central - PubMed

Affiliation: Key Laboratory of 3D Information Acquisition and Application of Ministry, Capital Normal University, Beijing 100048, China. lihanlun@126.com.

ABSTRACT
This paper describes an airborne high resolution four-camera multispectral system which mainly consists of four identical monochrome cameras equipped with four interchangeable bandpass filters. For this multispectral system, an automatic multispectral data composing method was proposed. The homography registration model was chosen, and the scale-invariant feature transform (SIFT) and random sample consensus (RANSAC) were used to generate matching points. For the difficult registration problem between visible band images and near-infrared band images in cases lacking manmade objects, we presented an effective method based on the structural characteristics of the system. Experiments show that our method can acquire high quality multispectral images and the band-to-band alignment error of the composed multiple spectral images is less than 2.5 pixels.

No MeSH data available.


The band images; (a) the infrared band; (b) the red band; (c) the green band; (d) the blue band.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4541943&req=5

sensors-15-17453-f004: The band images; (a) the infrared band; (b) the red band; (c) the green band; (d) the blue band.

Mentions: We select four synchronous images acquired by our multispectral system, which contain many artificial objects, as shown in Figure 4. The number of SIFT feature points of the four images respectively is 5097, 6481, 5911, and 6618. Because green vegetation has higher reflection levels in the infrared and green bands, and lower reflection levels in the red and blue bands, green bands are more similar to infrared bands in vegetation areas, and the registration between the green band and the other two visible bands is easier, so the green band is chosen as the reference band image, and the other three will be registered to it. The total initial matches respectively are 1446, 1139, and 485. For all initial matches, the coordinates of the key points in the reference band image are respectively (x01, y01), (x02, y02), …, (x0n, y0n) where the coordinates of the corresponding points in the input band image are respectively (xi1, yi1), (xi2, yi2), …, (xin, yin), and i = 1, 2, 3 represents the points in the other three bands. Letting dxi = x0 – xi, dyi = y0 – yi, dx3 and dy3 respectively indicates the matches’ coordinate displacements between the green band image and the infrared band image in X direction and Y direction. To represent distribution of dx3 and dy3, two histograms have been created, as shown in Figure 5. We can see that the number of matches reaches the top when dx3 is at −6.17; the number of matches reaches a maximum when dy3 is at −8.84. It can be estimated that (xrow, ycol) is approximately equal to (−6.17, −8.84) for the green band and infrared band pair. We can use the same method to calculate the displacements between the green band and other two bands. If flying height and flying attitude vary a little, the evaluated (−6.17, −8.84) can be used to register other images taken at other times. The correct matches respectively are 1278, 973, and 257, with the correct rate of 88%, 85% and 53%. The correct rate of initial matches between the green band and the infrared band is significantly lower than the correct rates between the green band and the other two bands. However, because of many manmade objects, all the correct rates are higher than 50%, so we can still use the RANSAC method to eliminate the error matches directly.


A Multispectral Image Creating Method for a New Airborne Four-Camera System with Different Bandpass Filters.

Li H, Zhang A, Hu S - Sensors (Basel) (2015)

The band images; (a) the infrared band; (b) the red band; (c) the green band; (d) the blue band.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4541943&req=5

sensors-15-17453-f004: The band images; (a) the infrared band; (b) the red band; (c) the green band; (d) the blue band.
Mentions: We select four synchronous images acquired by our multispectral system, which contain many artificial objects, as shown in Figure 4. The number of SIFT feature points of the four images respectively is 5097, 6481, 5911, and 6618. Because green vegetation has higher reflection levels in the infrared and green bands, and lower reflection levels in the red and blue bands, green bands are more similar to infrared bands in vegetation areas, and the registration between the green band and the other two visible bands is easier, so the green band is chosen as the reference band image, and the other three will be registered to it. The total initial matches respectively are 1446, 1139, and 485. For all initial matches, the coordinates of the key points in the reference band image are respectively (x01, y01), (x02, y02), …, (x0n, y0n) where the coordinates of the corresponding points in the input band image are respectively (xi1, yi1), (xi2, yi2), …, (xin, yin), and i = 1, 2, 3 represents the points in the other three bands. Letting dxi = x0 – xi, dyi = y0 – yi, dx3 and dy3 respectively indicates the matches’ coordinate displacements between the green band image and the infrared band image in X direction and Y direction. To represent distribution of dx3 and dy3, two histograms have been created, as shown in Figure 5. We can see that the number of matches reaches the top when dx3 is at −6.17; the number of matches reaches a maximum when dy3 is at −8.84. It can be estimated that (xrow, ycol) is approximately equal to (−6.17, −8.84) for the green band and infrared band pair. We can use the same method to calculate the displacements between the green band and other two bands. If flying height and flying attitude vary a little, the evaluated (−6.17, −8.84) can be used to register other images taken at other times. The correct matches respectively are 1278, 973, and 257, with the correct rate of 88%, 85% and 53%. The correct rate of initial matches between the green band and the infrared band is significantly lower than the correct rates between the green band and the other two bands. However, because of many manmade objects, all the correct rates are higher than 50%, so we can still use the RANSAC method to eliminate the error matches directly.

Bottom Line: For this multispectral system, an automatic multispectral data composing method was proposed.For the difficult registration problem between visible band images and near-infrared band images in cases lacking manmade objects, we presented an effective method based on the structural characteristics of the system.Experiments show that our method can acquire high quality multispectral images and the band-to-band alignment error of the composed multiple spectral images is less than 2.5 pixels.

View Article: PubMed Central - PubMed

Affiliation: Key Laboratory of 3D Information Acquisition and Application of Ministry, Capital Normal University, Beijing 100048, China. lihanlun@126.com.

ABSTRACT
This paper describes an airborne high resolution four-camera multispectral system which mainly consists of four identical monochrome cameras equipped with four interchangeable bandpass filters. For this multispectral system, an automatic multispectral data composing method was proposed. The homography registration model was chosen, and the scale-invariant feature transform (SIFT) and random sample consensus (RANSAC) were used to generate matching points. For the difficult registration problem between visible band images and near-infrared band images in cases lacking manmade objects, we presented an effective method based on the structural characteristics of the system. Experiments show that our method can acquire high quality multispectral images and the band-to-band alignment error of the composed multiple spectral images is less than 2.5 pixels.

No MeSH data available.