Remote Sensing Images Target Detection Research Based On Deep Learning And Model Of Self-evolution

The detection of ship target using large-scale,high-resolution visible light remote sensing images has always been a research hot spot in the field of remote sensing,figuring prominently in the military and civil fields.The detection of ship targets in a large range of high-resolution visible light remote sensing images is facing great challenges due to the complex and changeable background of ship targets,multi-scale,multi-class and multi-attitude intra-class differences of ship targets and the limitations of the detection model.Important approaches for large-scale ship target detection include ship target interpretation based on image information,such as synthetic aperture radar(SAR),thermal infrared sensor,optical sensor,etc.Therein,synthetic aperture radar(SAR)has been extensively applied in military,civil and other fields due to its capability of allweather and all-weather observation,as well as the ability to penetrate the obscuration of clouds and fog under certain band conditions.Therefore,SAR is widely used in military,civil and other fields.However,information can not be effectively extracted because of the dismal readability of SAR images,making it extremely important to conduct research on the target detection and recognition in SAR image.Traditionally,follow-up work need to be supported through manual extraction of SAR image features.In recent years,the role of deep learning(DL)in the field of SAR image target detection and recognition has received increasing attention of researchers due to its wide application in various industries.However,there are still some problems remain unsolved due to the imaging characteristics and shooting angle of SAR images.First of all,plenty of redundant information in the SAR image caused by the aerial view shooting will lead to smaller size compared with the background information.In the existing detection algorithms for small target ships,the detection accuracy cannot be greatly improved due to the loss of important information caused by feature extraction.Secondly,affected by water clutter,complex water background,sea fog and other changeable weather,the background redundancy information is prone to be more prominent than the ship target itself,resulting in false detection due to complex background.Thirdly,due to the condition of wide resolution of SAR images,the ship targets of the same category may have different sizes and abnormal aspect ratio under different resolution acquisition.These targets are prone to separate themselves from the range of receptive field applicable to the detection algorithm,and their edge information is easily ignored by the existing detection algorithm of convolutional neural network.Finally,the detection accuracy can be affected by the different ship sizes in SAR images and complex background in the context of the reasons above.An adaptive algorithm is required in the detection of the ship target,and its adaptability needs to be improved through an independent network framework.The above mentioned problems indicate the challengeable characteristic of ship detection.Therefore,the present study mainly focuses on the corresponding algorithms proposed to solve the existing problems in ship detection.(1)The one-stage object detector in the third chapter of this paper is proposed to solve the problem of missing detection and misdetection caused by many small-size ship targets in the complex background.In this work,a dense connection operation is added to the backbone feature extraction network to fuse the high-level and low-level features of SAR image containing ships,so as to reduce the loss of semantic information of smallsize targets.Then,the convolutional block attention module(CBAM)is inserted into the network in the feature-fused neck module.More characteristic information about different depths of input images can be expressed and the redundancy of information can be suppressed by integrating attention mechanism into the top-down Feature Pyramid Networks(FPN)and the bottom-up Path-Aggregation Network(PANet),thus producing a one-stage object detection model based on dense connection mechanism and attention mechanism.This model can not only balance the speed and precision of ship detection in practical tasks,but also improve the accuracy of network detection of ship targets and small-size ship targets in complex scenes.(2)In view of the target with abnormal aspect ratio in SAR image and the problem that some edge information is easy to be ignored,The fourth chapter of this study explores superior methods during the detection process.The self-attention module(Transformer mechanism)starting from the global feature is introduced according to the characteristics of the local optimal solution produced by the convolution neural network and the visual selection mechanism.The multi-head attention module in the self-attention mechanism is introduced into the one-stage object detection framework and fused with the convolutional neural network,so as to enable the transformer mechanism to combine the extraction of global information with the position sensitivity of the convolutional neural network.By imitating the working characteristics of biological visual selection mechanism,the attention operation from global to local during the operation of the model can be realized,thereby extracting the semantic information of ship targets more comprehensively and richly,and improve the detection precision of SAR images.(3)In combination with the contents of the previous two chapters,we get down to investigate how to make machines complete the unsupervised and autonomous learning like creatures,so that the ship targets with different characteristics can be more widely self-adaptive detection.With this mind,we examine the self-evolutionary processes of organisms and summary the relationship between need and evolution based on relevant theoretical knowledge.Accordingly,a need-based model of the structure of the selfsystem is proposed,which conceptualises and models the need and internal functioning mechanisms of the organism.In addition,we make efforts to combine the agent model of reinforcement learning with our self-system structure model,and update the reinforcement learning model according to the characteristics of need selection and evaluation mechanism in the self-model,so as to form a demand-based agent selfevolution model.Experiments are carried out to demonstrate the feasibility and validity of our model.(4)Combined with the previous research results,an adaptive object detection model is established according to the self-evolution model theory of agents.The stages of biological visual selective attention mechanism are introduced in different modules of the framework as follows: pre-attention,attention stage and prediction stage.In the preattention stage,the model automatically classifies the image size in the dataset according to the definition of the target size in the MS-COCO(Microsoft Common Objects in Context)dataset.On this basis,this model can be used to simulate the selective attention process of human vision to features when processing images in an autonomous and effective manner.In the attention stage,the model in chapter four is integrated into the above model.Corresponding feature fusion operation is carried out according to the different size of the target in the pre-attention stage.Finally,the detection results are predicted using mathematical methods in the prediction stage.The model has a certain degree of autonomy in adaptive detection of targets with different characteristics by modeling the self-evolution model theory and simulating the process of biological visual selection mechanism,thus improving the detection precision and flexibility of the model.The above-mentioned target detection models for SAR images have been tested on public remote sensing dataset,and their superiority has been verified by comparing with the most advanced target detection algorithms in the field of remote sensing.The experimental results demonstrate high detection precision of the models proposed in this study,as well as obvious advantages in real-time ship detection,which is conducive to a favourable balance between speed and precision.These improvements have potential significance for the application of real-time ship detection in specific fields.