New feature extraction algorithm for satellite image non-linear small objects

For satellite image non-linear small objects, a new feature extraction algorithm is proposed. This algorithm extracts features in a hierarchical structure. The features which represent the global shape property of the objects are extracted in inferior levels and the features which describe more local details of the objects are extracted in superior levels. In the first-level phase, the algorithm extracts the image binary entropy and the image normalized moment of inertia. In the second-level phase, based on wavelet transform as the detail micro tool, the algorithm extracts the Hu moments, the Zernike moments and the Fourier descriptors for all the child wave band images, and the features of each sub-wave band are respectively weighted according to their descriptive power. All the feature extractors are made invariant to translation, rotation and scale. The minimum Europe distance classification experimental results, the FCM and SVM recognition experimental results demonstrate that the algorithm can gradually describe the non-linear small objects in satellite images from global to local, from rough to fine. Compared with classical Hu moments, the Zernike moments and the Fourier descriptors, this algorithm is able to offer an available more competitive feature extractor for pattern recognition of satellite image non-linear small objects.