Research On Ship Target Detection Based On Deep Learning

China is rich in water transportation resources,with long and narrow mainland coastlines and island coastlines,as well as criss-crossing inland waterways.Ships play an important role in water traffic.Accurately identifying and locating ships can reduce the incidence of water traffic accidents and increase channel flow,which is of great significance for the intelligent research of ship detection and the construction of water transportation industry.With the improvement of computer performance and data volume,deep learning has gradually replaced traditional target detection algorithms in the field of target detection.In the field of ship target detection,the target detection method using deep learning can not only enrich the research results,but also have practical significance for the intelligentization of water transportation.Based on the method of deep learning,this paper studies two improved ship target detection algorithms.The main research work is as follows:(1)Study and analyze the status quo of target detection and ship target detection at home and abroad,and expound some basic theories of deep learning algorithms and several classic target detection network models.The data set used in this paper is introduced,and the robustness and generalization ability of the model can be improved by enhancing the diversity and quantity of the data set.The evaluation metrics related to the object detection field are introduced to make the experimental results in the paper more convincing.(2)Aiming at the shortcomings of the low detection accuracy of the SSD algorithm,especially for small objects,an improved SSD model is proposed.Replacing the backbone network,the ResNet50 network with better feature extraction capabilities is used to replace the original VGG16 network structure,and the hollow convolution is used in ResNet50 to replace the ordinary convolution,so that the backbone network has a larger receptive field.The multi-scale parallel receptive field module is introduced in the back layer structure of the backbone network to further increase the receptive field of the high-level feature map,and improve the feature extraction ability of the previous backbone network.Finally,the CBAM attention mechanism is introduced to enhance the semantic information on the feature map to improve the detection ability of the model.The ablation experiment and comparative experiment are set up in the experiment to verify whether the improved method is effective and the impact of the improvement on the accuracy of the model.The experimental results show that the improved algorithm has a significant improvement in accuracy and small target detection,and the improved SSD algorithm has a detection accuracy of 94.47%.(3)To further improve the accuracy of ship object detection,an improved YOLOv3 model is proposed.Firstly,the problems of the original YOLOv3 algorithm on the ship data set and the shortcomings of the network itself are analyzed.In response to these problems,corresponding improvements are made on the original YOLOv3 algorithm.The prior frame of the original YOLOv3 algorithm cannot cover the ship target very well.The K-means clustering algorithm is used to re-cluster the ship data set to generate a new prior frame suitable for the ship target.The SPP module is introduced in the network,so that the lower layer feature map can make full use of the semantic information of the shallow layer,and the expressive ability of the feature map is enhanced.By referring to the idea of residual structure,the SPP module is improved so that it can obtain more feature information.In the predictive regression part,the CIOU Loss function,which can more comprehensively measure the relationship between the real bounding box and the predicted bounding box,is used to calculate the localization loss.The ablation experiment and comparative experiment are set up in the experiment to verify the improved method and the impact of the improvement on the accuracy of the model.Experimental results show that the accuracy of the improved YOLO3 algorithm reaches 96%.