Research On Ship Detection Method Based On Data Set Subdivision And Band Screening

Ship detection at sea is an important part of the detection and management of maritime activities,and it is of great significance in military,people’s livelihood,economy and other fields.The traditional ship detection generally adopts the manual visual method,which has high cost and low efficiency.Some achievements have been made in ship target intelligent recognition methods based on satellite borne,airborne and UAV remote sensing images.However,the construction of ship data sets and the selection of effective bands for ship recognition need to be further studied.In order to achieve efficient and accurate identification of ship targets,this paper focuses on the improvement of detection accuracy of deep learning ships based on data set segmentation and band selection based on sample segmentation principle and variance analysis method.The main contents and results include:(1)Ship data set construction and attribute feature analysis based on UAV images.The UAV was used to acquire the image data of offshore vessels in bingjiawan,Huludao port,Xingcheng wharf and other places in Huludao,Liaoning Province.According to the vessel classification standard in the shipping industry standard of the people’s Republic of China(GB / t14-2011),the acquired vessels were classified into class 4vessels(passenger ships,passenger cargo ships,cargo ships,yachts),class 8 vessels(fishing ships,agricultural ships)and Class 0 vessels(other).The original data of the ship is preprocessed by rotation,clipping and widening,and the four characteristics of the ship image,including bow,material,deck and aspect ratio,are statistically analyzed.Among the No.4 ships,the bow features are mainly pointed bow,and the materials are all steel ships.The cabin or cargo hold accounts for a large proportion of the deck,and the ships are mostly slender;In class 8 ships,the bow shape is mainly divided into arc shape,pointed shape and flat bow,and the material is made of wood and steel.The cabin is small or there is no cabin,and there are many kinds of irregularly arranged fishing gear on the deck.The hull is relatively wide;In class 0 ships,the bow shape is mainly pointed and flat.The material is rubber.There is no cabin on the deck and the deck space is small.The length width ratio is about 2.3:1.(2)Vessel detection based on data set subdivision.Based on the original data set,the subdivision data set is established to realize the two-level and three-level subdivision of the data set.The neural network model is trained to compare the detection accuracy between different classification methods.The results show that the accuracy of the neural network model can be improved by further dividing the same class of ships into two or more subcategories(two-level subdivision)through certain specific characteristics,avoiding confusion and missing detection among various ships during detection,and the accuracy of ship detection can be increased from 90.41% to95.24%.(3)Vessel detection based on band filtering.The ship image is separated by single band image to study the influence of target background difference on the detection result.Some single band samples are randomly selected for the statistics of the absolute value of the gray difference between the target and the background,that is,the statistics of the target background difference.The statistical results are 26.83912 in B band,15.18459 in G band and 53.22767 in R band.The accuracy of ship target detection under the conditions of R,G and B bands is obtained,which is 80.83% in B band 81.67%in G band and 89.17% in R band.The r-band image has the largest target background difference and the highest detection accuracy.The research results show that the parameters such as the length and width of the ship can be used as the effective parameters for ship detection,and the ship detection method based on the two-level subdivision of the original data set according to the length to width ratio can effectively avoid the detection confusion and missing detection among ships,and the accuracy of ship detection can be effectively improved;The band filtering method is beneficial to the shallow feature extraction of ship targets,and the band images with large difference between target and background are more suitable for the detection of ship targets.