Design And Implementation Of Maritime Object Detection System Based On Deep Learning

With the adventure of the 21st century,land resources are gradually saturated in the process of people’s development and utilization,and the strategic significance of the ocean to the country continues to increase.Under this situation,it is more and more important to ensure the safety of country’s waters and build a object detection system that can effectively detect objects at sea.The thesis firstly improves the algorithm based on NanoDet,an anchor-free target detection algorithm.The SE module(Squeeze-and-Excitation Module)is introduced into the ShuffleNet v2 of NanoDet’s backbone network;The FPN-PAN module is introduced at the FPN(Feature Pyramid Networks)layer,replacing the previous non-convolutional minimal PAN(Path Aggregation Network)module,which improves the quality of the acquired multi-scale features;Combined with the characteristics of maritime targets,a new metric GBEP is introduced to improve the loss function.After that,experiments were carried out on the improved NanoDet algorithm.The experiments show that the improved NanoDet object detection algorithm has a certain improvement in detection accuracy and speed in the test data set.Secondly,based on the improved NanoDet algorithm,the thesis introduces the concept of key frame,combined with the context relationship of video data in the temporal domain,the video object detection algorithm based on improved NanoDet is implemented by methods such as key frame feature fusion and non-key frame feature sparse transfer,which further improves the speed of object detection.Finally,on the basis of the video target detection algorithm,the thesis analyzes the requirements of the system from the functional requirements and non-functional requirements,clarifies the overall goals to be achieved by the system,completes the overall architecture design of the system and the design and implementation of each module of the system,and implements system testing.