Abstract: A feature pyramid network (FPN) is proposed, which resolves low detection accuracy of underwater small target, due to lack of feature information and inaccurate positioning. Firstly, convolution, horizontal and vertical pooling are used to extract static and dynamic context information of feature maps, which is named coordinative nonlocal attention module and plugged into upsampling process of FPN. Secondly, parallel dilate convolution and efficient channel attention network (ECANet) are used to capture multi-scale space and channel feature information, which is named trident feature enhancement module and plugged into channel adjustment process of FPN. Finally, the linear regression loss gain coefficient is introduced into the regression loss function of the Faster R-CNN algorithm, to increase the penalty for the multi-scale target regression offset and improve the positioning accuracy. The experimental results show that, compared with the baseline Faster R-CNN algorithm, the proposed algorithm improves the detection accuracy by 2.8%, 2.2% and 2.5% in the dataset provided by the 2020 National Underwater Target Detection Contest, PASCAL VOC dataset and the MS COCO dataset, respectively, which proves the effectiveness of the method.