ADAPTIVE WEIGHTED MULTI-TASK SPARSE REPRESENTATION CLASSIFICATION IN SAR IMAGE RECOGNITION

In this paper, a novel multi-task sparse representation (MSR) of the monogenic signal is proposed in order to overcome the misclassification caused by heterogeneity of three components of the monogenic signal. In recent years, the monogenic signal has been applied into the field of SAR image recognition due to its capability of capturing the broad spectral information with maximal spatial localization. The monogenic signal can be decomposed into three components (local amplitude, local phase and local orientation) at different scales. The components are concatenated to three component-specific features and then fed into a MSR classification framework. However, the heterogeneity of the three component-specific features makes it difficult to make decisions by simply counting the accumulated error in multi-task sparse representation classification. To solve this problem, a multi-task learning model based on Fisher discrimination criteria is designed and Fisher score is presented to measure the discriminative ability of three types of component-specific feature in different classes. The final decision is made by weighted accumulated reconstruction error. Experiment results prove the effectiveness of adaptive weighted MSR classification method of monogenic signal.