Research On Space Target Few-shot Classification And Small Target Detection Algorithm Based On Deep Learning

In recent years,space exploration technology has advanced at a breakneck pace,the number of different types of satellites in space has continued to increase,and the amount of space information brought by this has also ushered in a period of rapid growth,which brings great challenges to the development of space target surveillance and situational awareness.Artificial intelligence technology,on the other hand,has advanced quickly in recent years.Deep learning methods represented by image detection and recognition and information retrieval have brought dawn to the continued development of spatial information technology.How may these image data be used to classify and detect space targets is a recent research hotspot.However,there are still some difficulties in using deep learning to empower space exploration technology: First,because the space targets run too fast in space,the angle of the space imaging equipment is relatively single,which brings about the problem of too little image information for some non-key space targets,especially non-cooperative targets.Traditional image feature extraction and classification methods are difficult to accurately classify objects with few samples.Second,because the distance between some imaging devices and the space target is too far,the imaging quality of the target is poor,and the number of pixels occupied by the target is too small.Conventional detection methods are difficult to find specified targets for tracking and detection in the space background.In order to solve the above problems,this thesis studies and implements a deep learning-based space targets few-shot classification and small target detection algorithm.To begin,this thesis proposes a metric meta-learning-based few-shot classification technique for space targets.The algorithm is applied using an end-to-end deep neural network model,which combines an attention-based feature fusion module and a confidence-based prototype network for few-shot classification of space targets.To address the problem of the model’s low generalization ability to new target samples due to the small sample size,the method of feature fusion of sample support set and query set is adopted to strengthen the information about the target in the few samples,and improve the generalization ability of the model to new samples.Secondly,by calculating the confidence score of feature classification,it is used to distinguish the features of different types of targets,so as to create the prototype of the classification network,and effectively identify the targets of different categories in the high-dimensional space.The experimental results show that the few-shot classification algorithm proposed in this thesis outperforms the traditional classification algorithm and current mainstream few-shot classification algorithms.On the BUAA dataset,when the number of samples is 5,the accuracy is 3.79% higher than the latest MCT(Meta-Learned Confidence Transductive)few-shot classification algorithm.Next,this thesis proposes a space small target detection algorithm based on a single-stage detection network,which solves the problem of stellar interference in target detection by connecting the target detection model with the deep classification model.For the problem that the target size in the space is too small,a convolution enhancement module that integrates multiple convolution kernels and multiple dilated convolution is used to strengthen the ontology features and adjacent spatial features of small targets,and is paired with a single-stage multi-scale target detection network,which increases small-space target detection ability.The experimental results indicate that the algorithm described in this thesis can considerably increase the detection accuracy of small space targets,in addition to improving the detection effect of various scales on natural images.Compared with the current mainstream YOLOv5(You Only Look Once version 5)algorithm,only by adding17% model parameters,the detection accuracy can be improved by 1.82%.The few-shot classification of space targets and the small target detection algorithm researched and implemented in this thesis can be used in various tasks of space situational awareness,such as target positioning,tracking,and monitoring the surrounding environment to prevent accidents that may occur at any time,which is important in terms of both theory and practice for the future growth of space exploration.