Research On Small Target Detection Method Based On Scale Generation Network In Remote Sensing Images

Object detection is the core technology of computer vision,which is widely used in intelligent video surveillance,autopilot,aviation monitoring and other fields.At the same time,with the development of remote sensing technology,the amount of remote sensing data is increasing rapidly.Therefore,remote sensing image object detection has gradually become a research hotspot,it refers to the design of a target detector,which can effectively identify and locate the target of interest in remote sensing images.However,there are many serious problems in this field.First of all,because the objects in the image are extensive and messy,resulting in the problem of complex background interference during detection.Secondly,there are many and dense small targets in the remote sensing image,so there is lack of sufficient small target information for recognition.Finally,the target scale varies widely,often including only a dozen pixels of small targets such as cars,while containing hundreds of pixels of large-scale targets such as basketball court.How to design a robust detector to effectively solve these problems needs to be discussed.In view of the above-mentioned difficulties,this paper proposes a remote sensing image target detection network based on global spatial semantics and local context,and uses deep learning-based methods to carry out algorithm and model research on target detection problems in remote sensing images.The main research work is as follows:In view of the complex background of remote sensing image from the perspective of aerial photography,which seriously interferes with the detection accuracy,and the problem that it is difficult to extract global features by traditional methods,this paper proposes a multi-scale perception network,which constructs a multi-level attention network in the feature extraction network to extract more useful global context features.By using the attention mechanism to extract the global semantic information of the image,the network pays more attention to the target area with strong information expression and suppresses the interference of complex background.In view of the problem that small targets in remote sensing images are often clustered and lack of sufficient information,and the traditional detection methods often cause the loss of small target information through multiple down-sampling operations,this paper proposes a small target information enhancement network.It learns the context relationship between spatial local regions of the image through Long Short-Term Memory Network,and explores the semantic correlation between the region around the target and the target itself,providing auxiliary detection clues for the target with insufficient information by remembering the useful features of the region around.Aiming at the problem of multi-scale target detection in remote sensing images,traditional methods usually set up many groups of candidate frames with different proportions and sizes in advance as the screening reference frame,but this method will produce a lot of useless redundant frames,which is inefficient.Therefore,a self-adaptive anchor generation network is proposed in this paper.By using the rich semantic feature information of the image to be tested,the network is guided to generate unconstrained center position and size,and scale self-adaptive candidate frames are generated.It can not only effectively detect targets with large scale transformation,but also has strong adaptability to scale changes of different datasets in remote sensing field.