Complex Gaussian–Bayesian Online Dictionary Learning for SAR Target Recognition with Limited Labeled Samples
| Autor: | Lanqi Wang, Qian Wu, Qingsen Han, Biao Hou, Licheng Jiao |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Synthetic aperture radar
General Computer Science business.industry Computer science Bayesian probability ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION General Engineering 020206 networking & telecommunications Context (language use) Pattern recognition 02 engineering and technology Target acquisition Complex normal distribution ComputingMethodologies_PATTERNRECOGNITION Discriminative model 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing General Materials Science Artificial intelligence business |
| Zdroj: | IEEE Access. 7:120626-120637 |
| ISSN: | 2169-3536 |
| DOI: | 10.1109/access.2019.2935164 |
| Popis: | For solving the problem of limited synthetic aperture radar (SAR) labeled samples, an initial SAR target recognition algorithm based on complex Gaussian-Bayesian online dictionary learning is here presented. The amplitude and phase information of SAR images is an important discriminator for target recognition, which derives significant statistical distribution-based target recognition. First, to better fit the SAR images and to reduce the computational complexity, a complex Gaussian distribution (CGD) model in the context of dictionary learning was established to model SAR images. Second, as the discriminative dictionary can be learned in conjunction with modeling the distribution characteristics of SAR images, a discriminative dictionary of the distributed model had to be learned. Finally, to solve the problem of limited labeled samples and the time consumption of the existing algorithms, the semi-supervised online dictionary learning method was used to add the training samples to update the dictionary. The moving and stationary target acquisition and recognition (MSTAR) dataset was used to complete the experiment, and then, several comparison methods were used to ensure fairness. Experimental results revealed that the proposed algorithm was better than the compared algorithms consistently in the case of different-sized training samples. The proposed method can reach an accuracy of 94.52% when using 20% training samples which is much higher than the comparison algorithms. Moreover, the proposed method is 0.5% higher than the second-best method when using the whole training samples. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|