Remote sensing target detection based on visual saliency guidance and classifier fusion

It has much realistic significance to rapidly detect targets for large-field remote sensing images, especially within limited computation resources. Inspired by the cognitive characteristic and structure of selective attention in the human visual system, a novel model of target detection based on visual attention mechanism was presented, which combined bottom-up visual saliency and top-down interpretation of salient-region. In this proposed model, the detection was divided into three serial stages: pre-attention, attention and post-attention. Specifically, an adaptive and morphological strategy was employed to generate saliency map for selecting salient regions from entire scene rapidly. Finally, to discriminate the task-related targets from other similar salient-objects, a top-down and task-dependent method based on the classifier fusion technique was introduced. The comparative experiments over a ship detection task validate the effectiveness of the proposed model. Simultaneously, in such a manner of hierarchical and concentrative computation, the system resources can be reasonably distributed.