Abstract: An improved loop closure detection algorithm based on scene salient regions is proposed for VSLAM (visual simultaneous localization and mapping) loop closures without consistent path and similar camera view. Firstly, a scene salient region based loop closure detection architecture is structured together with its probabilistic model. Then a key frame selecting method which combines feature tracking rate and RGB histogram matching is proposed to reduce the detection information redundancy. Next, a pre-matched scene selecting method based on inverted index is presented, which greatly improves the real-time performance of loop closure detection. Finally, according to the geometric property of the salient region in this paper, a geometric matching probability in observation is introduced to improve the algorithm's ability of distinguishing perceptual aliasing scenes. The comparative experiments show that the proposed algorithm gets a better recall ratio compared with FAB-MAP 2.0 in a condition of high precision and good real-time performance, when facing loop closures without consistent path and similar camera view.