An artificial target detection method combining a polarimetric feature extractor with deep convolutional neural networks
Rui Sun, Xiaobing Sun, Feinan Chen, Hao Pan, Qiang Song
Abstract
To improve the success rate of target detection, a fusion method that comprises a polarimetric feature extractor and a deep convolutional neural network (CNN) with consecutive small (2 × 2) convolutions (or, CSC-CNN for simplicity) is proposed. First, we theoretically analyse the dispersion characteristics on a target surface based on the concept of information entropy, and it is concluded that the dispersion measure can be selected as a relevant feature for polarimetric images. Then, a polarimetric feature extractor is introduced for conveniently calculating dispersion measures with fewer prior parameters in outdoor measurements. Finally, a CSC-CNN is adopted for subsequent target detection with small scale training samples. The experimental results indicate that the proposed fusion method demonstrates great potential for reducing the detection error rate, which is less than that of the traditional method used for comparison.