Fast Detection Method For Marine Objects Optimizations Based On Deep Learning

It is of great significance to monitor the distribution of ships in key sea areas for the development of marine economy,the supervision of maritime traffic,the protection of maritime safety and the construction of national defense.However,ship object detection is faced with many problems,such as extreme environment interference in complex sea area,complex satellite optical image data,and small ship object size.Therefore,it is an urgent problem to detect the specific ship’s category and location quickly and accurately.In recent years,with the wide application of deep learning technology in the field of computer vision,deep learning provides an important technical basis for ship object detection.In this paper,the method of marine object detection based on deep learning is studied.Aiming at the problems of complex extreme environment interference,large amount of optical image data and small object size,an efficient object detection framework with modified dense connections for small objects optimizations(DEOD)is proposed.The innovative work of this paper mainly includes:1.In order to improve the detection speed of marine objects,from the point of view of optimizing the structure of deep convolutional neural network and according to the characteristics of small object detection of ships,this paper compresses the model of object detection framework and designs a cross layer concatenating dense network structure with parameter optimization,which uses the dense connection between multiple convolutional layers to the context.The multi-channel features are concatenated and recombined to reuse the cross-layer features,so as to compress the number of parameters,reduce the computational complexity and improve the detection speed.The experimental results show that the network structure can achieve about 30%compression ratio of the object detection framework compared to the benchmark framework and increase the speed by about 22.6%while maintaining the detection accuracy.2.In order to improve the detection accuracy of marine objects,the feature extraction network in this paper integrates the residual structure based on the dense connection structure,and the object detection framework designed in this paper adopts a hybrid structure of dense and residual structure,so it can map the cross-layer features identically to alleviate the gradient vanishing caused by the deepening of the network,and further improve the detection accuracy.The experimental results show that the proposed framework based on the network structure is improved by 9.2%mAP compared to the benchmark framework.3.In order to further improve the detection accuracy of small ship objects,this paper proposes a multi-layer feature fusion method of the same scale based on dense structure.On the basis of extracting multi-layer regional features at different scales,the method uses the cross-layer concatenated dense network structure to fuse the position features and category features of the same scale between the dense network convolutional layers to provide richer semantic information for the prediction layer as information supplement.Besides,the method further explores the small object features of the shallow layer and optimizes the detection accuracy of small objects.The experimental results show that the detection accuracy of DEOD with this method improves the detection accuracy of small-sized ships by 26%AP compared to the benchmark framework.