Image recognition is not very effective in the water environment due to multiple factors, such as high scattering and high scattering in the water column. This is why the relevant parameters in the Faster R-CNN network model need to adjust continuously to improve the effectiveness of water detection. The control variable method adjusts the program's learning rate by tuning the network model's parameters. Then, the number of training rounds is adjusted according to the loss function of each round, and finally, we can get the number of matches with the minimum loss function. Based on the experimental results on the dataset, it is shown that the proposed method not only selects the learning rate with the best detection results but also has the strongest robustness and achieves a 96%-99% recognition rate for passenger ships, cargo ships, warships, and bridges compared with other learning rates. Experiments show that the Faster R-CNN network model detects water targets with significant results, and the best network model learning rate parameter is 6×10-3.

Abstract
1. Introduction
2. Related Work
    2.1. Convolutional Layer
    2.2. Pooling Layer
    2.3. Fully Connected Layer
    2.4. Classification Layer
    2.5. Faster R-CNN Networks
3. Experiments
    3.1. Experimental Dataset
    3.2. Training
4. Experimental results and Discussion
    4.1. Experimental results
    4.2. Experimental Discussion
5. Conclusion
References

• ChaoYu Lu(College of Naval Architecture and Port Engineering, Shandong Jiaotong University)
• FengGuang Jia(College of Naval Architecture and Port Engineering, Shandong Jiaotong University) Corresponding author
• Li Min Yu(College of Naval Architecture and Port Engineering, Shandong Jiaotong University)