A deep learning based underwater object detection model is proposed to address the problem of low detection accuracy of multiple underwater targets, which is due to the blurred underwater environments and the fusion between underwater targets and background. To improve the receptive field and subtract more effective topmost feature, the YOLOv5 detector is modified by adding deformable convolution and dilated convolution. Further, Swin Transformer is introduced into the backbone of the modified YOLOv5 to obtain more information of small underwater objects. Instead of coupled detection head, a decoupled head is designed in the modified YOLOv5 to reduce the interference between data by predicting classification, regression and confidence information separately. Combined with an underwater dataset, the improved underwater multi-objective detection model is verified. The detection results show that the accuracy of the model for small underwater target detection has been significantly improved.