A comparison of neural network and classical texture analysis

Textures in high resolution radar images may be characterized in terms of their single point statistics and correlation properties. For example, in synthetic aperture radar images, textured regions may be modelled reasonably well by correlated K distributions. For some image analysis techniques, such as image segmentation, it is desirable to be able to classify such textures in a manner which is as close to optimum as possible. The performances of a number of texture classification schemes are compared with the maximum likelihood classification. The schemes which are considered fall into the three categories of autocorrelation function fitting, density estimation and neural network classification. The performances are assessed by classifying simulated textures composed of either Gaussian or K distributed single point statistics.