| In recent years,object detection has a lot of applications in the field of remote sensing,such as environmental monitoring,disaster detection,rescue assistance and so on.In these applications,object detection based on deep convolutional neural networks(CNNS)plays a key role.With the development of convolutional neural networks,breakthroughs have been made in remote sensing object detection based on Faster RCNN,YOLO and SSD algorithms.But these methods rely heavily on large amounts of annotated data.Getting large amounts of manually annotated data often takes a lot of time and effort,and even in some cases,it is unrealistic to get large amounts of annotated data.On the other hand,models with a small number of labeled samples are prone to serious overfitting and generalization problems.These factors limit the application of the model to larger categories.In order to solve the above problems,some scholars try to use the method of few-shot object detection to solve the problem of scarce training data of remote sensing object detection.Few-shot object detection is a new research direction of computer vision with the development of few-shot learning and object detection technology.In recent years,researchers have proposed many few-shot object detection methods,such as image-based metric learning,model structure-based,and meta-learning,but these methods are mainly aimed at natural scene images.As remote sensing image objects are compared with natural scene images,their orientation and size distribution are diverse,and their backgrounds are relatively complex.Therefore,it is difficult for existing few-shot object detection methods to obtain ideal detection results when applied directly to remote sensing images.First of all,on the basis of algorithm study and experimental verification on existing few-shot object detection,For the characteristics of arbitrary direction and difficult to detect in remote sensing targets,this paper proposed few-shot object detection algorithm on remote sensing images based on multi-directional feature enhancement.The existing few-shot object detection algorithm mainly targets the natural scene image with better target direction consistency,does not consider the problem of different target direction,this paper puts forward the adaptive feature fusion module based on the dual pipelines few-shot object detection framework,and selects the corresponding direction category attention characteristic according to the change of target direction,so as to obtain more targeted category auxiliary information and improve the model detection ability.At the same time,in view of the positive and negative sample imbalance in training,a positive and negative sample sampling method based on difficult mining is proposed,and the effectiveness of the algorithm is verified by a large number of experiments.Secondly,aiming at the problem of diverse size distribution of remote sensing images,this paper proposes a few-shot remote sensing target detection algorithm based on multi-layer feature fusion and dual-attention enhancement from two aspects of model and loss function.From the perspective of the model,on the one hand,a multi-layer feature fusion method is proposed based on the feature pyramid network,which improves the expression ability of features at different levels by combining the deep semantic information with the shallow location information.On the other hand,the dual attention enhancement module is constructed,which can reduce the interference of background information to target through channel and space dual attention enhancement,and enhance the ability of network feature extraction.From the point of view of Loss function,an improved Focal Loss function was proposed to improve the model’s attention to small-scale targets.Finally,based on the proposed method and the actual requirements,a few-shot object detection on remote sensing images prototype system is constructed.User-defined image category and number,model learning,object detection and other functions are realized... |
PDF Full Text Request