Abstract: To improve the detection accuracy for armored targets in complex environment, a salient target detection algorithm was proposed. The low-level features and multi-scale pooling semantic features constrained by visual attention mechanism were respectively obtained by visual attention mechanism and joint pyramid upsampling module. Then the aggregation strategy was used to fuse, so as to improve the ability of target representation in low contrast or occlusion conditions. The experimental results show that the proposed algorithm obtains good detection results for multi-scale targets in complex scenes, the precision, recall rate and mean average precision (mAP) are 72.2%, 71.4% and 77.1%, respectively, which can meet the practical application requirements.