Target Overlapped Scenarios Based On Improved Ship Target Detection Method For RFBnet

Ship is an important carrier of maritime activities.The ship target detection based on computer vision has been applied to the actual ship management system,therefore,the detection of ships near shore and on the sea surface has different application values in various fields.At present,in such images,the ship targets under the condition of multiple targets are easy to be blocked by multiple targets,causing problems such as missed detection of small targets and misclassification.How to improve the accuracy and speed of detection and meet the needs of Marine security in practical applications is an urgent problem to be solved.Traditional ship target detection algorithm is divided into three types,such as target detection based on statistics,based on knowledge of the target detection,target detection based on the model,the algorithm requires human to extract target feature.With the maturity of deep learning theory,algorithms based on deep learning have greater advantages than traditional target detection algorithms.Therefore,in the detection of ship targets,the ship targets under the condition of multiple targets are easy to be blocked by multiple targets,resulting in missed detection,classification error and other problems of ship targets,and the study is carried out by using the deep learning method.Firstly,a natural image target detection method based on improved RFBNET(I-RFBNET)is proposed.Firstly,the pooled feature fusion module(PFF)and deconvolution feature fusion module(DFF)are used for feature fusion,and six new effective feature layers are formed.Secondly,a step long volume method was proposed to extract the interested region information of the feature unit in the original image,and a dilate convolutions block(DB)integrated with the attention mechanism was designed to perform feature fusion with the new first three effective feature layers.Then the focused classification loss function is introduced to solve the unbalanced distribution of positive and negative samples in the training process.The experimental results show that the improved algorithm has a good detection effect,especially for small targets in the case of multiple targets.The average accuracy is 96.26%,which is4.74% higher than the previous improved algorithm.Second,Small targets for multi-objective situation around easily and multi-objective overlap or color overlap with the environment,by improving and optimizing the I-RFBnet network model,using the K-means++ clustering algorithm of adaptive learning prior box wide high proportion,K-means++ clustering algorithm can provide good initial conditions for the follow-up frame regression,and add the repulsive force Loss function,put forward a kind of Improved RFBnet and Repulsion Loss CNN(IRRL-CNN)network model,The model further improves the accuracy of detection by using repulsive force loss.Various comparative experiments also verify that the I-RFBNET algorithm based on the repulsive loss function has a better detection effect for overlapping targets.