Abstract: Feature representation is a key component to salient object detection.However,the features extracted by existing methods lack capability of discrimination for salient object detection.Hence,multi-scale context features extraction mechanism and attention recurrent mechanism are proposed to solve this problem.Specifically,the multi-scale context features extraction mechanism uses atrous convolution,which can expand receptive fields of high-level convolution features to obtain rich context sematic features,and adopts a vector aggregation strategy to fuse these features.In order to enhance the discriminative power of fused features,the proposed method adopts a attention mechanism to distinguish the importance of each pixel by adaptively increasing the weight of convolution features.The attention mechanism can focus on salient regions while suppressing the interference information in the background.Furthermore,a recurrent network is presented to gradually refine the convolution features in spatial position,and adjust salient regions to generate accurate saliency maps.The proposed method is compared with eight state-of-the-art methods on five benchmark datasets.Experimental results show that the method can generate more accurate and complete saliency maps,and achieves good performance in both mean absolute error and maximum F-measure.