Research On Typical Object Detection Method In Optical Remote Sensing Image

With the development of remote sensing satellite sensor technology,high resolution and multisource data that acquired from satellite,aviation and drone platforms is no longer a difficult problem.Driven by so many databases,the detection tasks that could not be completed due to data limitation in the past can be gradually realized at this stage,which promotes the continuous technology research and exploration in the field of object detection in order to meet the increasing requirements.Especially for urban areas which is closely related to human beings or military hotspot areas which are related to national defense,for the precise detection of vehicles,aircraft and other object,no matter when it is used for urban road planning,traffic management,optimization and enhance convenient index of people's lives,and providing information support for military object positioning and tactical object striking,they all have important theoretical studies significance and practical value.The flowchart of object detection usually consists of three stages: image segmentation,object feature extraction and object detection.For the stage of object region segmentation in image,most algorithms mainly use pixel and region based segmentation methods in early phase,they utilize texture information differences between classes to acquire the segmentation results,which is suitable for large remote sensing areas segmentation,and lacking the consideration for object.Therefore,people have proposed many superpixel segmentation algorithms focus on object,by constructing various constraints and replacing pixels with superpixel blocks as processing units,which greatly improve image segmentation accuracy and processing efficiency,but it still exists a low maching between segmentation boundary pixels and real object boundary pixels,which caused the position of the superpixel center is shifted and further leading to the inaccuracy problem of object patch extraction and subsequent feature expression.In addition,image segmentation method is usually applied on the ideal image without noise,but in the actual data processing,the influence of noise for the image can not be ignored.For object feature extraction stage,the traditional model features have limited information representation ability,and they no longer meet the detection requirements with complex urban background.In recent years,deep learning technology has shown strong feature learning ability and has been widely used,but it is not universal about the number of training samples,network structure design,parameter setting and so on.Moreover,most of deep learning based object detection methods adopted well-established training databases,but they are not suitable for specific tasks,so it is necessary to select and establish training sample sets according to specific detection tasks,but traditional manual or random based sample selection methods both exist some insufficient,which influence the selection of the optimal training samples,and further restricting feature expression and influencing detection precision.In image detection stage,most methods use different classifiers to classify the object and background,to obtain the final detection results,but they are still subjected to use single-source data information or single feature,how to utilize so abundant data from different platforms and sensors to mine the different features of single-source data or the complementarity and redundancy of multisource data,and it still has big promotion space about the detection precision.So this paper proposed novel ideas and technologies for obtaining high precision of object detection have important research significance through analyzing the problems are faced in different object detection stages.Firstly,in order to extract the complete object structure from the image,for object region segmentation,a multiple local information constraints based superpixel segmentation method is proposed,which adding multiple local information constraints on the segmentation boundary pixels besides of color and space constraints,the matching between image segmentation boundary pixels and real object boundary pixels is improved,and segmentation breakage is reduced,so as to ensure the accuracy of extracting object patches.For the disadvantageous effect on image segmentation results because of noise,a 3D sparse code based image denoising algorithm is proposed,sparse dictionary and sparse coefficients are updated iteratively by K-SVD algorithm using three-dimensionan spatialspectral feature blocks extracted from images as input.Finally,sparse recovery model is used to output high quality images.The experimental evaluation consists of two parts,the proposed superpixel segmentation method is compared with other two advanced superpixel segmentation algorithms,the proposed image denoising algorithm is compared with the traditional sparse representation based image denoising algorithm.Two groups of experiment are shown that the proposed two algorithms both obtain the best performance through visual sense and quantitative evaluation,to provide an important fundamental guarantee for subsequent feature extraction and detection.Secondly,based on the results of superpixel segmentation,the object patch can be extracted based on superpixel center(patch is an image block containing object or background).For object patch feature extraction stage,due to the weakness of traditional feature representation ability and the deficiency of training sample selection method cause a low detection precision,an object detection based on CNN feature extraction method is proposed,which utilizes the strong feature learning and representation ability from convolutional neural network,to obtain deep semantic information of object samples,then adopting training sample automatic selection method,and ensuring the distinction of the within-class and the discriminability of the inter-class to construct the optimal training sample set.Meanwhile,in order to solve object structure is incomplete after patch clipping and the difference of feature extraction caused by different object directions,an object main direction automatic rotation algorithm is adopted to unify object direction.The experimental evaluation includes the comparison with training sample set selection method,the comparison with several traditional feature and advanced CNN feature based methods.The proposed object detection based on CNN feature extraction method obtains a best performance.Finally,detection accuracy can be significantly improved through the above mentioned that CNN feature based object detection method,but it exists that the detection precision is low when detection recall rate is high,so for the restrict of single feature detection results using single feature,an optimization method of object detection results for single source data is proposed.Firstly,in order to make full use of the spectral feature of data,a local tensor discriminant analysis based image classification algorithm is proposed,which achieves the high accuracy of image classification results by combining spectral-spatial features with tensor discriminant feature extraction,then an environmental elements constraints based inference method is proposed,which fuses the image classification results with the existing detection results,to remove false objects and improve detection precision.In addition,in regard to multisource image object detection with small object size and few training samples,a spatial-spectral feature collboration based object detection in multisource data algorithm is proposed,which utilizes the advantage of superpixel segmentation and abundant multisource spatial-spectral features,and the spatial and local pooling spectral features are extracted form multisource object patches,to significantly improve the detection precision in single-source data.