Scene recognition is a key problem in mobile robot topological navigation. For unknown environments

a natural scene recognition approach based on visual local salient regions is presented. Firstly

a feedback saliency detection model (FSDM) is presented to carry out bottom up scene image analysis. Then

according to the salient positions

automatic scale selection is realized based on fractal dimension to build the local salient regions with appropriate size. Those salient regions are represented by 3 invariant features of gradient orientation

moment and canonical hue. They are used for scene recognition in terms of their match ratio. Experiments show that FSDM can obtain higher accuracy. The scene recognition experiments in both indoor and outdoor environments show that the approach has high stability and tolerance compared to the method based on global appearance when scale and viewpoint etc changed. The accuracy of recognition for static scene is higher.