The Fractal Feature Extraction Technique Of Vehicle Targets In High Resolution SAR Imagery

With the rapid development of synthetic aperture radar(SAR) and the diversification of its carrier, it is capable to monitor the battlefield in real time. To better interpret the future battlefield, it is urgent to detect and classify ground vehicle targets(such as tanks, armored vehicles, trucks, missile launchers, etc.), especially when the resolution of SAR imagery approaching to sub-decimeter level. As widely reported, the false alarm rate as well as the classification accuracy is sharply reliant on feature extraction, a core step in SAR imagery interpretation. To deal with the false alarms, this paper first analyzes the scattering phenomenology difference between target and background. Then, fractal theory is exploited to quantitatively measure this difference, thereby the vehicle targets and clutters can be discriminated easily.The thesis first reviews the basic concepts, physical properties, research status and development trend of fractal theory. We know that the back scattering in SAR imaging is complicated and the vehicle targets pixels have obvious gap. But it is not effective to describe above phenomena using the common Euclidean geometry. To handle this problem, the thesis introduces the fractal features, such as fractal dimension feature, extended fractal feature and lacunarity. Moreover, the generation of diversiform lacunarity features via fractal theory is presented.We first present a duplicate variance lacunarity feature based on the existing one. The box mass is defined as the variance of the pixels amplitude in the box and that is the first variance. Then we can get the duplicated variance as a new lacunarity feature by calculating the variance of above box mass. Multiple comparative experiments demonstrated that it is preferable and stable to use the duplicated variance to discriminate the vehicle targets.According to the significant difference between vehicle targets and clusters in SAR imagery, the thesis proposes a multidimensional layered lacunarity feature vector based on fractal theory. This feature can eliminate the false alarm by quantitatively describing the vehicle targets' profile lacunarity and its pixels' degree of irregularity. In the process of calculating this feature vector, we first linear transform the test slices in order to unite the dynamic range of pixels' gray level. Secondly, we proliferates layers by center to the surrounding and sets each layer's pixel amplitude variance as a component of multidimensional layered lacunarity feature vector. Finally, the fuzzy C-means clustering algorithm is used to process the feature vector to get a cluster membership function which we can use to discriminate the vehicle targets from natural landscapes.Simulation SAR images, MSTAR database and domestic airborne SAR image data are adopted to test and comparatively analyze the capability of all the features in this thesis.