Research On Ship Target Recognition In Optical Remote Sensing Image Based On Fusion Of Multifeature In Deep Learning

As an important means of transportation at sea and the main target of attack in wartime,the research on ships is of great significance for both military and civilian application.Compared with synthetic aperture radar(SAR)image,optical remote sensing image has higher spatial resolution and lower recognition cost.Therefore,this paper mainly studies ship target recognition based on optical remote sensing image.In addition,compared with other target recognition tasks,there are many kinds of ship images,and image acquisition is vulnerable to illumination,occlusion,marine environment and other factors,which makes the task of ship recognition more challenging.As one of the applications of deep learning in image processing,Convolutional Neural Network(CNN)has gradually become a research hotspot in the field of pattern recognition.Compared with traditional algorithms,CNN has the advantage that the acquired high-level feature is more abstract and separable,but it has some shortcomings in the application of ship target reco Blition.Therefore,based on the complementarity between CNN and traditional algorithms,combined with the characteristics of ship image,this paper proposes two different algorithms of ship target recognition.The main contributions of this paper are as follows:Firstly,because convolutional neural network does not achieve global rotation invariance and cannot capture low-level features completely,which are particularly important for ship image classification,a ship recognition method based on convolutional neural network and multi-scale rotation invariant features is proposed.On the basis of using CNN to obtain more high-level features,this method employ feature fusion strategy to introduce multi-scale rotation invariant features,which can make full use of spatial structure,local texture,different orientations and other characteristics of ship images.The proposed algorithm achieves excellent classification performance on two ship data sets,which has been confirmed to be superior to some state-of-the-art methods.Secondly,a multi-channel CNN fusion framework based on multi-domain features is proposed to solve the problem that ship images are vulnerable to clouds,sea surface conditions,imaging sensor parameters that result in blurring ship contours and affect feature extraction.The first part is to extract transform domain features,including Gabor features,CLBP features,multi-order amplitude inverse transform features,multi-order phase inverse transform features,etc.The second part is the learning process of multi-branch CNN.The feature maps are composed of multi-channel images as inputs of CNN to get features that are more abstract.The third part uses decision fusion strategy to integrate the classification results of different branches:into the final classification results.Decision fusion strategy can integrate these advantages from each branch to achieve better experimental results.We verify the classification performance of the proposed algorithm on three ship data sets.The experimental results demonstrate that the proposed method can achieve better results.