A Linear-Feature Based Method for the Quality Assessment of Remotely Sensed Images
Quality Assessment by Linear Features
Linear features can be detected by the so called edge detectors. Since all edge detectors are more or less based on the computation of gradients (grey value differences between adjacent pixels), our quality assessment method by linear features is to some extend also related to the previous method based on the variance estimation. Because variance represents also the degree of variation in grey value, but more of global variation. While edge detectors looks only for local differences.
It is well known that gradient is very sensitive to noise, and we have seen from the previous demonstration dangerous the noise could be, therefore it might seem not to be a wise idea to use linear features for judging the quality. But by using a very small trick we can totally avoid the influence from noise. This is exactly the reason why we propose this method. We know that random noises are un-correlated and can not cause long linear features, therefore simply by limiting the minimum length of the linear features, we can get rid of the problem associated with noises. Also the problem of the definition of noises is also easily solved by choosing a proper size for the mask of the edge detectors. It is a common practice in the edge detectors. It is a common practice in the edge detecting to filter out or to suppress high frequency signals which would disturb the detection. For example, by selecting a proper size of the LoG or the Canny operator mentioned below, signals above a certain frequency can be smoothed out. Here we can use this ability to solve the problem according our purpose of application, information above a certain frequency should be treated as noise, then by choosing a proper size for the edge detector, all information higher then that frequency can be effectively suppressed.
There are many well known edge detector methods. Any one could be used for our purpose. Important is that after edge detection, the detected edge elements have to be linked together to form linear features. Since our goal is to compare image quality relatively, it makes no difference which edge detector and which edge linking method are used. As long as all the images are judged by the same methods, relative comparison of their quality unchanged.
In our experiment we use the Zero-Crossing method proposed by Marr et al. (1980) for detecting the edge elements. Since their original operator reacts also to very weak edges, we have set a proper threshold to suppress very weak edges. The linking is achieved by simply following the flagged pixels of zero-crossing. Two indices can be derived for judging the image quality. One is the total number of linear features in the image. The second is the average length of the linear features. The same two SPOT images as used in the previous chapter are used to demonstrate this new method.
Table 2. shows the quality assessment results. Similarly like Tab 1, image column 1 and 2 are for the original images and the third column is for the artificially corrupted 299303L. We can see that in the absence of noise. 299303R has more linear features than 299303L. According to our definition of image quality, 299303R is better than 299303L. Therefore the linear features method shows the same results as the variance method.
Tab. 2 Quality assessment by linear features for the SPOT images
| image |
299303L |
299303R |
299303L
corrupted |
| Band |
B/W |
G |
R |
IR |
B/W |
G |
R |
IR |
B/W |
G |
R |
IR |
#linear
features |
11995 |
6708 |
3104 |
3627 |
14068 |
7985 |
4371 |
3085 |
4613 |
2194 |
1384 |
1821 |
Average
length |
9.0 |
8.8 |
9.0 |
8.9 |
9.1 |
9.3 |
9.0 |
9.1 |
9.0 |
8.8 |
9.0 |
8.9 |
From the third column we can see that the linear features method still indicates that 299303R is of better quality than the corrupted 299303L. Our method is not affected by the artificially added noise . Remember that while using the variance method for judging the image quality, if the noised strength is not considered, the corrupted 299303L will be wrongly judged as better. Here again we see the advantage of linear feature based method. There is no need to do any elaborated estimation of the noise variance.
Conclusion
With increasing commercialization of satellite images, objective quality assessment of remotely sensed images becomes more important. Since the payment or the acceptance of ordered images depends on the image quality.
After analyzing the quality assessment method based on the estimation of the signal strength, we proposed a very simple linear-feature based method for assessment of the image quality. The new method is much easier to implement because it does not have to do the very complex estimation of the noises. The computation is much less than the traditional variance method. The only weakness of this method is that it gives no absolute measurement of the image quality. The index is only relative. Which means one need a "good" image as reference first and compare other images to this one to get the assessment of their quality. But for the practice, it is not a problem to select a proper reference image.
References
- Canny, J. 1986 "A computation approach to edge detection." IEEE Trans. PAMI, vol. 8, pp. 679-697.
- Kraus, K.; Mihail, E. M. 1972: "Linear Least-Squares Interpolation." Photogrammetric Engineering. Vol. 38, pp. 1016-1029.
- Marr, D.; Hildreth, E. 1980: "Theory of Edge Detection." Proceedings Royal Society London., Vol. B207. pp. 187-217.
- Meer, P.; Jolion, J. M.; Rosenfeld, A. 1990: " A Fast Parallel Algorithm for Blind Estimation of Noise Variance." IEEE Trans. PAMI, Vol. 12. pp. 216-223.
