Currently, underwater organism detection tasks are generally carried out with mobile devices. In the field of underwater target detection, it is difficult to deploy the detection model to mobile devices due to the large parameters. The GCP-YOLOv7 based on YOLOv7 was proposed in this paper to cope with problems such as the difficulty of detecting underwater targets in mobile devices and the limitation of hardware resources. Firstly, the ELAN module in the Neck part was lightened by the GhostNetV2 module to reduce model parameters and computation. Secondly, to solve the problems of missing features and lower accuracy that the lightened network may have in collecting feature information, we added the CA attention module after the improved ELAN module to prevent feature missing. Finally, we adopted a 50% pruning rate to prune the overall improved model to reduce the model parameters and computation. Compared with the YOLOv7, the parameters and computation of the GCP-YOLOv7 are reduced by 4.25 and 4.08 times, respectively, and the accuracy is improved by 2.8%.