Research On Ship Target Recognition Technology Based On Deep Transfer Learning

China has a vast territorial sea and abundant marine resources.As the main body of marine activities,the realization of accurate identification of ship targets plays a vital role in the construction of a maritime power.At present,the ship target data collection methods are diverse,mainly including spaceborne,airborne,unmanned aerial vehicle,shipborne and shore-based,etc.There are certain differences between the image data obtained by different collection methods;The ship target depth recognition network obtained from one platform’s image data cannot be directly applied to other platforms.This is due to the various differences in observation angle and image resolution between platforms.Therefore,this paper focuses on the research of ship target recognition based on deep transfer learning,and the main tasks are as follows:(1)An improved CDAN deep transfer learning method based on human-in-the-loop is proposed.First,in the target domain data pseudo-label prediction and generation link of the Conditional Adversarial Domain Adaptation(Conditional Adversarial Domain Adaptation,CDAN)method,the human-in-loop mechanism is introduced,and the label correction tool software developed by us is used to realize the pseudo-label prediction and generation of the target domain data.Efficient and convenient manual review and correction of tags effectively solves the problem of deep network performance degradation caused by false tag prediction errors;secondly,in the feature extraction module of the CDAN method,an aggregation transformation method is added to achieve feature extraction capabilities,which is an effective way to improve the field Adaptive performance provides support;finally,in the domain adaptive module of the CDAN method,a new feature modeling layer is added to realize the modeling of domain-specific features,thereby effectively solving the problem of negative migration impact caused by domain-specific features on the migration process.(2)Through a large number of comparative experiments and analysis,the effectiveness of this method is verified.First,on the public data set Office-31,compared with CDAN method,the method in this paper has an average recognition accuracy improvement of 1.5%;second,on the public data set Office-Home,the method in this paper has an average recognition accuracy improvement compared to the CDAN method.Third,on the self-built ship target data set,compared with the CDAN method,the method in this paper has an average recognition accuracy improvement of 5.1%.Finally,on both the public and self-built data sets,the method in this paper is compared with There are six other deep transfer learning methods including the CDAN method.The method in this paper achieves the best overall deep transfer network performance.(3)After completing the above theoretical research and experimental verification,comprehensively using Python,Py Qt and other development tools to design and implement a ship target recognition prototype system based on deep transfer learning.The system has functions such as tag correction tools and ship target recognition from the front view angle to the squint view angle.