| Summary: | 碩士 === 國立宜蘭大學 === 綠色科技學程碩士在職專班 === 105 === Target recognition from hyperspectral image is detection an object from cover or camouflage on the image, such as a special material, mineral or military targets and so on. The main process of the target recognition is as follows: First, select one (or multiple) targets from the image, then calculate the suppressed background, and then use the target detection method to obtain the similar target spectral signatures in the image. Although many target detection methods have been proposed, various methods of target detection have different detection precision because of their different assumptions.
In this thesis, the first proposed method is to improve the existing target detection method based on the spectral angle mapper. The method includes the following steps: firstly, define the maximum angle parameter of the spectral angle mapper for searching similarity objects in the hyperspectral image, that is, to confirm the meaningful spectral signature set; then compute the mean value of the spectral signature set, Thirdly, re-search the entire hyperspectral image to identify targets with similar or identical spectral characteristics.
The second improved method is to sort the original detection results, remove the spectral signature set with the highest similarity in the ranking, then re-calculate the background information with other spectral signatures, and search the whole hyperspectral image again, with similar spectral signatures of the target.
Receiver Operating Characteristic (ROC) Curve is used to evaluate the performance of different algorithms. The results show that the proposed two methods can effectively improve the detection accuracy of Adaptive Coherence Estimator (ACE) and Orthogonal Subspace Projection (OSP). According to the characteristics of the hyperspectral image, the first method can improve the accuracy of target detection by 1% ~ 5%. The second method can improve the target detection accuracy by 2% ~ 6%.
|
|---|
