Research On Ship Target Recognition Technology Based On Deep Convolutional Neural Network

China has a vast territorial sea and abundant marine minerals.The coastal border situation is very complicated.As the main body of the ocean,the realization of the automatic identification of ship targets is an important guarantee for safeguarding the national territorial sea security and strengthening the construction of a maritime power.In recent years,a large number of on-board,UAV based,satellite-based and shore-based video surveillance equipment have been installed,and a large number of video image data have been collected.However,most of the ships still use traditional manual methods for recognition,and the degree of intelligence is very low,which can not meet the requirements of automatic recognition of ship targets.In the process of ship target recognition by traditional methods,the recognition performance is poor due to environmental changes such as light,waves,heavy rain and salt spray.Therefore,how to achieve robust automatic ship target recognition is a hot and difficult problem to be solved.Thanks to the continuous deepening of artificial intelligence researches,especially the breakthrough and successful application of deep learning in the field of image target recognition,this essay studies the application of deep learning in automatic ship target recognition.The main work is as follows:(1)Firstly,the domestic and international research status of ship target recognition is reviewed,and the current RFCN deep network with good comprehensive performance is introduced;(2)Aiming at the difficulty of automatic ship target recognition in visible video images captured by marine environment,a ship target recognition method based on the improved RFCN network is proposed,and three modules are improved and optimized in details.Firstly,in the feature extraction module,a combination of aggregation conversion and multiple-scaled feature fusion is designed.By methods of adding improved branches of the same topological structure,combining the low-level large-scale location information and high-level small-scale semantic information,the feature extraction ability has been effectively improved.Secondly,in the target detection module,a combination of bi-linear interpolation,sensitive detection of cascade positions and flexible non-maximum suppression is designed;while fine pooling,high-quality positive samples,flexible attenuation confidence and other mechanisms can effectively enhance the detection ability of small sized and overlapping ship targets.Finally,in the RPN module,for ship target recognition applications,clustering the ratio of the width and height of ship target anchor frame can effectively improve the ship's target detection ability;(3)At first,the experiment of super-parameter value selection is designed and carried out;then,the comparison experiment between this method and RFCN method is implemented:on the VOC 2007 data inspection set,this method has achieved a 2.42%improvement in mAP.On the self-built ship target dataset,the proposed method achieves a mAP increase of 3.51%.However,in terms of recognition speed,this method reduces the 1FPS compared with the RFCN method,so the method is applicable to applications where computing resources are not limited;(4)Using Java,JavaScript,Python,Caffe,Redis,RocketMQ and other development languages and middleware tools,a visual prototype platform for ship target recognition based on deep convolution neural network is designed and realized.