Research On Approximate Nearest Neighbor Query Method For Key Ship Target Recognition

The recognition of the key ship targets at sea is one of the key directions in target detection field.The traditional method is based on the image processing of remote sensing to complete the task,but the existing monitoring technology of remote sensing image is often affected by various equipment performance(such as image resolution),natural factors(such as climate,sea state)and human factors(such as confidentiality),therefore the recognition performance is limited.In addition,the traditional method only relies on the single information of remote sensing images,which is one-sided,and is also a major obstacle to the practical application of remote sensing target recognition technology.In order to consider comprehensively all aspects of information and improve the recognition accuracy,as an effective complement to rely solely on the remote sensing image technology,based on a variety of Internet text and image information such as public knowledge.By integrating them organically,a more intelligent and effective recognition system of the key ship targets at sea is designed.Based on the Internet resources,combined with the focus crawler technology,we crawl the corresponding ship-themed images and other multimedia information,and study the approximate nearest neighbor search method for the recognition of key ship targets.To reduce the quantization error of approximate nearest neighbor search method caused by hard coding in the process of large-scale image retrieval,we improve the product quantization method and present a method of soft encoding based on barycentric coordinates,reducing the quantization errors to make the result expression closer to the actual original data.The main research works in this paper are below:(1)For the social network platform,news media website and other data sources,we propose a four-layer image-oriented focused web crawler model based on the web crawler technology,and crawl based on the specific theme.It improves the data filtering effect for the subsequent approximate nearest neighbor image retrieval method provides an image dataset source.(2)Put forward an algorithm of image retrieval based on barycentric coordinate product quantization(BCPQ),which encodes high-dimensional feature space into sparse representation and uses the concept of barycenter to set quantization encoding mode to reduce the quantization error effectively.(3)Based on the above models and algorithms,we design and implement an approximate nearest neighbor search system for key ship targets retrieval recognition,and describe the design details of each function module.The performance of the proposed model and algorithm is verified in the corresponding chapters.The experimental results show that the proposed model,algorithm and the designed and implemented system are effective,and can be applied to the field of military and civil ship target recognition.