The field of underwater target detection has very important research value in many research fields, such as marine environment, resource exploration, marine biological research and so on. Underwater target detection still has the following challenges: (1) The underwater target has high similarity with the background and low identification; (2) Generally, the volume of the underwater target to be detected is very small and difficult to find; (3) The underwater acquisition image is fuzzy and has heterogeneous noise, which requires targeted image processing and data enhancement technology; (4) Camera jitter will cause changes in the background area and affect the accuracy of target detection methods based on background modeling. In addition, there are many occlusions of underwater targets, which have a certain impact on feature extraction. Therefore, this paper proposes an underwater target detection framework based on YOLO v7, which is composed of supervised feature learning technology. Based on the new model scaling technology, a new Cross Stage Merge is designed to reduce the impact of scaling on the calculation amount and accuracy. By extending the efficient aggregation network, a new framework of E-ELAN is proposed. Finally, the accuracy rate of YOLO v7 at 30 frames per second known in GPU 2080TI exceeded YOLO v5 and YOLO v4, reaching 59.3% AP. The final experimental results are based on the training method from scratch, without the use of pre-training weights.