To address the problems of small targets detection with few available features and weak texture information in the context of complex sea conditions in multiple scenarios, and to improve the environmental perception capability of unmanned surface vehicles (USV), we propose a typical small targets detection method using attention mechanism and multi-scale features. Firstly, the global prior information of the target is fused in the deep layers of the network using atrous spatial pyramid pooling module. Secondly, the shallow spatial location and deep semantic information features of the target are adaptively enhanced by the attention fusion module to improve the feature representation capability of the network. Finally, the high performance target detection is achieved through multi-scale feature fusion. We construct a typical surface small target dataset, and the method is evaluated by experiments of surface small target detection under real sea conditions based on USV. Experimental results show that the proposed method in the NVIDIA platform reaches 17 FPS, which can accurately identify small target on the water surface. Compared with the original FPN algorithm, the mIoU is improved by 7. 58% , and the average detection accuracy is improved by 11. 41% to 82. 36% .