SAR ATR by Decision Fusion of Multiple Random Convolution Features

As the appearance of the target is sensitive to its aspect angle in one Synthetic Aperture Radar (SAR) image, the classification accuracy for target recognition task can be significantly improved through pose estimation when there are only limited training samples. A pose estimation algorithm is first proposed in this paper based on second moments of region of interest, which is enhanced based on an improved multi-scale fractal feature and detected by adaptive threshold segmentation. To compensate for pose estimation errors, two sets of random convolution features are extracted and concatenated using randomly generated kernels with a predefined set of sizes from the original and rotated images, respectively. Extreme learning machine (ELM) is used as base classifier, and an ensemble of ELMs is then trained based on feature resampling. Decision fusion rule based on the confidence of each base classifier is presented to combine these two classification results. MSTAR dataset is finally used to verify the effectiveness of the proposed algorithm in two standard and extended operating conditions, respectively.