Airport Aircraft Detection Based On Deep Environmental Context Modeling In Remote Sensing Image

With the development and improvement of modern remote sensing technology,remote sensing image object identification technology has been widely used in national defense,economy and people's livelihood.As an important carrier of aviation economy and military activities,aircraft has always been the focus of object detection.Especially for the automatic identification of airport aircraft,it can achieve the dynamic supervision of aircraft entering and leaving ports to build an intelligent control system,analyze the situation of non-cooperative airports to obtain war information,which has profound practical application value.Remote sensing airport images are large and complex,and there are various objects with similar apparent characteristics to aircraft.Moreover,the imaging process is affected by many factors,and the image characteristics of the object are prone to significant changes.Thus,the FAST searching ability and anti-interference ability of the detection algorithm are required.This paper focuses on the research of aircraft detection of optical remote sensing airport image,analyzes the local environment and discrimination characteristics of airport aircraft.In this paper,we focus on three key technologies of the complex airport scene hierarchical aircraft candidate region extraction,the deep environment context modeling based on features fusion,and the candidate region identification based on local environment constraints,and proposes an effective target detection method for airport aircraft.The main research work of this paper is summarized as follows:Firstly,a hierarchical candidate region extraction method based on spatial information is proposed to solve the problem of small proportion of targets in large remote sensing images,redundant background information,diverse non-target areas of the airport with inconsistent feature attributes,which makes it difficult to eliminate the problem effectively by single means.This method analyzes the characteristics of aircraft spatial distribution in airports,and adopts the reverse extraction method of eliminating non-object regions.According to the different texture feature attributes of non-object region,a method based on integral map was designed to remove the gentle texture region by constructing the gradient map and a series of morphological treatments.And a method based on two-dimensional cumulative scanning was proposed to eliminate dense texture region,which is combined with row and row scanning and marker location judgment.This method can realize the rapid extraction of aircraft candidate region and screen out a lot of invalid information.Secondly,aiming at the problem of high complexity of remote sensing airport scene and insufficient stability of object apparent information,a deep modeling method based on fusion of local context features and gradient features is designed.We analyze the difference between the object and the distractor,and select the context information as the expression.Based on MLPH descriptor,key points with aircraft local context information were extracted,and gauss weighting method was adopted to enhance the central feature of the key points,and the local context feature image was constructed.Then the multi-scale local context HOG pyramid is constructed by gaining calculation of the weight of the feature image and the gradient amplitude,effectively combining the local context feature and the gradient feature,so as to realize the depth information fusion modeling.The local context feature map is constructed,and the local context HOG pyramid is generated based on the gradient feature to realize deep information fusion modeling.By combining complementary information,the ability and stability of feature expression is enhanced.Then,a candidate region identification method based on the local information constraint is proposed.In complex scenes,false alarm is easy to be generated when identifying objects,due to insufficient description of features.In this paper,the local context HOG feature pyramid is adopted to identify candidate regions,by introducing local region environment information as constraints,to effectively overcome the problem of false detection caused by insufficient description of features.Compared with object identification based on a single feature,this method significantly improves the object discrimination ability and effectively suppresses the false alarm.What's more,to address the challenge of aircraft rotation variations,a method of prediction orientation of suspected target direction based on circular sweep filtering is designed,according to the information of aircraft geometry structure and parking arrangement.The problems of the large computation and low time efficiency are overcome,which are caused by full orientation rotation detection and mixed model detection.On the basis of the above three key technologies,this paper constructs a set of remote sensing airport aircraft detection system software.Through the rapid extraction of the potential region of the object,the fine identification is focused on the effective region,and based on the rich information level of the object features,the accurate detection of the airport aircraft is realized.At the same time,this paper set up a civilian and military-civilian airport images multiclass homemade dataset,with the support of this dataset,a multi-angle comparative experiment was carried out.Experiments show that the method proposed in this paper has good processing ability for large-scale remote sensing airport image,strong ability to resist complex scene interference,can realize accurate aircraft detection,and has high timeliness.