Abstract: Visual salient region detection is one of the key technologies for machine vision. There have been still many problems in salient region detection, e.g., the quality of the detection results is not high, and the whole salient region cannot be highlighted effectively, etc. Moreover, the existing salient region detection algorithms cannot effectively handle the salient region appearing in the edge of image, since they are easily affected by complex background and the color or texture of the object itself. Aiming at the shortcoming that the saliency detection model is not suitable for complex background and low contrast environment and the detection effect of saliency target located at the edge of image is poor, this paper proposes a saliency detection model based on multi-layer graph and compactness. Firstly, the image is divided into super pixels and a multi-layer graph is constructed by combining the high-level information of image block layer and cluster layer, which can effectively detect different scale images and obtain uniform salient regions. And then, both foreground and background seed nodes are extracted according to the compactness assumption. The sorting scores based on foreground and background seed nodes, respectively, are calculated by means popular algorithms via multi-layer graph model. Furthermore, the above two sorting scores are combined to get the saliency map, which is further filtered to obtain the smooth foreground and background regions and the final saliency map. The proposed multi-layer graph introduces the image block layer and the clustering layer and combines the local and global information so that it can reduce the sensitivity of seeds, obtain the uniform saliency region, and effectively detect targets of different sizes. Moreover, by means of the compactness calculation model, a more reliable seed point can be extracted to obtain more accurate results. Experimental results on MSRA-1000 database and ECSSD database show that the proposed algorithm can obtain higher accuracy and recall rate than nine popular methods.