Research On The Augmentation Method Of Ship Target Data In High-resolution Visible Light Remote Sensing Image

The rich and diverse feature information of high-resolution visible light remote sensing images has promoted the development of remote sensing image target detection technology.As an important carrier for marine traffic and marine resource detection,the rapid and accurate detection of targets on the sea surface is of great significance in the military and civilian fields.In order to meet the ever-increasing accuracy requirements of ship inspection,the current inspection methods have gradually transitioned from the traditional visual saliency threshold segmentation to data-driven inspection methods represented by deep learning.In order to meet the real-time processing application requirements of mobile terminals with limited resources such as storage and energy consumption,lightweight deep learning models are gradually used in ship inspection tasks because they can greatly reduce the amount of calculation.Among them,sparse Mobile Net V2 is a method for remote sensing on-orbit processing accumulated by the research team through previous experiments and research,which can effectively achieve fast and high-precision detection of ships under calm seas.However,due to the lack of ship targets and positive and negative sample classes in its training set,the problem of balance causes the detection model to easily fall into overfitting in complex cloud interference scenarios.Data augmentation is a technique to obtain more data through a certain transformation of limited training data.It can solve the over-fitting problem of ship target detection to a certain extent from the perspective of original data.This thesis proposes a data augmentation method(Space Association Data Augmentation,SADA)that combines sample space and feature space.It enriches the data set from the perspective of sample data transformation and feature space fitting to realize the augmentation and supplement of ship images.The specific research contents are as follows:(1)For the design of sample space data augmentation methods,which mainly rely on expert prior knowledge and image characteristics,through reference to common augmentation methods of natural scene image data sets taken on the ground,an experimental study was carried out to determine the strategy of ship sample space augmentation.Solve the shortcomings of the lack of research foundation of the ship target data augmentation method in the field of satellite remote sensing image analysis.First,by analyzing the characteristics of the ship image,the operation with the potential to reduce the difference between the training set and the real scene is selected from the sample space method,and the related parameters of the ship image are set in conjunction with the characteristics of the ship image in this thesis.Then,the original ship data is expanded by the same multiple through the selected method,and the recall rate and false alarm rate of the training set expanded by various methods on the sparse Moible Net V2 network are compared to determine the expansion of the ship target data based on the sample space in this thesis.Method specific operation method.According to the experimental results,the sample space augmentation method in this thesis can effectively improve the ship target detection recall rate,but at the same time brings a significant increase in the false alarm rate and reduces the model's ability to distinguish negative samples.(2)Aiming at the problem that the sample space augmentation method is not rich in target features due to data format restrictions and raises the detection false alarm rate,the generative adversarial network is used to augment the ship target data in the feature space.Aiming at the problem of the lack of clarity of the remote convolutional Generative Adversarial Networks(DCGAN)to generate remote sensing ship images,a remote sensing ship image is generated from the feature space fitting angle,and a deep convolution ship target that can further enrich the data set is proposed Generate an adversarial network(Deep Convolutional Ship Generative Adversarial Networks,DCSGAN),and verify its accuracy improvement effect through the sparse Moible Net V2 network.According to the experimental results,the feature space augmentation method in this thesis can improve the ship target detection recall rate while avoiding a significant increase in the false alarm rate,but its improvement effect on the recall rate is slightly insufficient compared with the sample space method.(3)Aiming at the insufficiency of image diversity in sample space and feature space augmentation methods and the contradiction between recall rate and false alarm rate,this thesis proposes a spatial joint data augmentation method for high-resolution visible light remote sensing ship targets.First,the original ship image is combined with the candidate data set obtained by parallel expansion of the sample space and feature space methods,and then combined according to the ratio determined by the comparative experiment to form the spatial joint augmentation data set.The performance verification and evaluation experiments carried out using the sparse Moible Net V2 network show that: compared with the single spatial data augmentation accuracy improvement method,the spatial joint data augmentation method in this thesis can significantly improve the recall rate while avoiding a significant increase in the false alarm rate,effectively Alleviated the conflict between the detection recall rate and the false alarm rate.(4)In view of the problem that the evaluation of target detection accuracy cannot comprehensively measure the effect of data augmentation,this thesis conducts a quantitative evaluation study on the quality of ship images generated by the DCSGAN network.Drawing on common evaluation indicators in the field of natural images,the Peak Signal to Noise Ratio(PSNR)and Structural Similarity Index(SSIM)are selected to measure the similarity in energy and structure of the ship image generated in this thesis and the original image.Solve the lack of research related to the lack of data augmentation evaluation in the field of remote sensing.Aiming at the adaptability of the feature extraction network Inception V3,which is a common evaluation index for feature similarity,to the effect of remote sensing ship image feature extraction in this thesis,an evaluation index Fre?chet Mobile Net Distance(FMD)for feature extraction with the sparse Mobile Net V2 network is proposed to measure the feature similarity between the generated image and the original image in this thesis.The experimental results show that the ship image generated in this thesis has less pixel distortion and higher structure and feature similarity compared with the real image.