Research On Underwater Target Detection System Based On Cascade R-CNN

With the continuous progress of science and technology,human’s ability to exploit the ocean has been greatly improved.The marine development environment is harsher than the land.In marine surveys or marine aquaculture and other activities,if relying on traditional manual methods,there will be not only high costs,but also high safety risks.As a hot research technology in the field of artificial intelligence,underwater target detection technology can play a huge role in marine exploration,marine protection and marine fishery.Designing an underwater target detection system applied to the marine environment will greatly improve the capability of marine development.The main research contents of this paper are as follows:Firstly,according to the current development status and performance indicators of underwater target detection,the function and performance requirements of the underwater target detection system are analyzed in detail.According to the analysis results,each module of the underwater target detection system is designed,and the whole detection process of the system is designed.Secondly,in view of the problems of color distortion in underwater images,a variety of traditional underwater image enhancement algoritluns are used.According to the shortcomings of traditional underwater image enhancement algorithms,such as the need for manual parameter adjustment and the lack of universality,an underwater image enhancement algorithm based on generative adversarial networks is designed.According to the underwater image evaluation index,the performance of these two types of underwater image enhancement algorithms is compared and evaluated,and the appropriate underwater image enhancement algorithm is selected according to the performance comparison and evaluation results.Then,the existing deep learning algorithm model is analyzed,and the Cascade R-CNN algorithm with better detection performance is selected as the underwater target detection algorithm.The algorithm uses ResNeXt101 as the basic network for image feature extraction,and uses deformable convolution and self-optimization The network model is optimized by three optimization schemes of thought data cleaning algorithm and random weight average algorithm.At the same time,according to the performance indicators of the target detection algorithm,the performance of the optimized Cascade R-CNN algorithm is evaluated from three perspectives:the individual evaluation of the optimization scheme,the overall evaluation of the optimization scheme,and the overall evaluation under different data sets.The evaluation results of the optimized Cascade R-CNN algorithm are compared with other underwater target detection algorithm models,and the feasibility of using the optimized Cascade R-CNN algorithm in underwater target detection is demonstrated.Finally,the experimental verification of the hardware platform of the underwater target detection system based on Cascade R-CNN is realized,and the TensorRT acceleration engine is used to accelerate the system hardware platform to improve the detection speed of the underwater target detection system under the GPU.At the same time,two hardware platforms,a deep learning graphics workstation and an embedded edge device,were selected to test the optimized Cascade R-CNN algorithm on the hardware platform.By comparing the optimized Cascade R-CNN algorithm model with other underwater target detection algorithm models The performance data under the hardware platform demonstrates the practical feasibility of the underwater target detection system under the hardware platform.