Research On Warship Detection And Recognition Based On Deep Learning

Warship detection and identification is the basis and premise of the realization of the realization of safety at sea,it can provide a strong protection and support for the country.Because traditional warship identification needs a lot of manual interpretation,and it is difficult to ensure the accuracy and reliability of warship identification,therefore,it is necessary to introduce new technologies and methods to ensure high efficiency of ship identification and reduce manpower consumption at the same time.With the advent of the era of big data,the development of deep learning in images provides new technologies for warship target detection and recognition.From the perspective of image technology,the dissertation designs a warship detection model based on the chaotic whale optimization algorithm and a warship semantic segmentation model based on CyclicNet to detect and recognize warships.Specific research contents are as follows:(1)Aiming at the problem that the non-maximum suppression algorithm greediness selects candidate boxes with high confidence,a method combining chaotic whale optimization algorithm with non-maximum suppression(CW-NMS)is proposed,which takes the post-processing algorithm as a combinatorial optimization problem.The chaotic search method is discretized into an initialized combination scheme,and the appropriate fitness function is designed considering the accuracy and recall rate.Under the guidance of fitness function,the local optimal combination is obtained,and the combination strategy is constantly updated by whale optimization algorithm.In addition,the difference set area method is proposed to optimize the combined results.Experiments are carried out on PASCAL VOC2012 and Warship target detection datasets.Compared with other common algorithms,CW-NMS improves the detection performance of the post-processing algorithm and makes the detection results robust.(2)A new convolutional neural network with a randomly connected clique network can improve the efficiency of network features and obtain more accurate classification results.In order to improve the transmission of semantic information,a new encoder-decoder based full convolutional network,CyclicNet,is proposed in this dissertation by introducing the randomly connected clique network in the CliqueNet network.Experiments are carried out on the street view datasets Cam Vid and Warship semantic segment datasets.The experimental results show that,compared with the current advanced semantic segmentation architecture,CyclicNet not only maximizes the information flow to achieve feature refinement,but also improves the image segmentation results.