Building Detection From High-resolution Remote Sensing Images Based On Visual Perception

In recent years,with the development of machine vision,cognitive neurology and artificial intelligence,building detection from high-resolution remote sensing image has been transformed into a target recognition problem.With simulating the process of cerebral vision perception,to optimizate or reconstruct the algorithm framework of building detection from high-resolution remote sensing image,is undoubtedly a meaningful attemption for the current high-resolution remote sensing image analysis technology.Building detection is a typical application of artificial intelligence and image processing technology in high-resolution remote sensing images,and which has a very broad application prospect in urban planning,urban management,population estimation and prediction,smart city construction and so on.Through the combination of visual perception technology and pattern recognition technology,it can overcome the large scale and complexity problems in high-resolution images.Based on the visual attention mechanism,multi-scale feature fusion,local invariant feature description,this paper proposes a hierarchical framework of building detection from high-resolution remote sensing image.The framework focuses from artificial area to building area,then to building contour dectection.In each layer,different feature extraction methods and analysis strategies are proposed for different recognition tasks,through continuously perceiving and focusing the scene and activating the underlying features from bottom to up.The main research can be summarized as the following aspects:(1)Artificial area automatic extraction method with multi-scale and multi-feature fused bsed on visual perceptionBased on the theory of feature integration in visual attention mechanism,the paper proposes a multi-scale and multi-feature fusion method for automatic artificial area extraction.Firstly,this method extracts the geometric features such as corner points,straight lines,right angles,and visual significant features such as edge features,main direction and orthogonal significance,local brightness,local contrast.Then the geometrical radiation intensity images and multi-scale visual significant images are obtained by adaptive fusion strategy,and further the artificial area saliency image is obtained.The paper proposes a filtration mechanism between the same geometric feature and also different geometric features,thus it can drive the artificial area detection better.And also the papar constructs the main direction and orthogonal significant(OOS,Orientational and Orthogonal Saliency)index,it can take account of both the unidirectional road and orthogonal buliding.The experimental results show that the method can not only extract the lage scale artificial area,but also detect the scattered and independent artificial objects easily,and also the method is of robustness and better adaptability.Meanwhile,it is more accurate and the missing rate is fairly lower than MRF and QGA.(2)low-dimensional description of hybrid features of building patches in high-resolution remote sensing imageBy simulating the structure of function columns of human's brain,the papar proposes a low-dimensional description method for the hybrid features of building patches in high-resolution remote sensing image.Based on the opponent-colors theory,the description method extracts and intergrates the brightness,similarity and color feature from the LAB color space,and realizes the low-dimensional description for the hybrid features of the building patches.The visual perception system is simulated by the binarization method whose dimension is obviously lower than that of representative local feature description method,such as SIFT and SURF.Compared with other representative local feature description method,the experiments strongly indicate that the proposed method is of better computing efficiency and more accurate classification performance.For patches of non-building area,patches of building contour and patches of building roof,the classification accuracy is near or above than 80%,which is acceptable to carry out building area and contour detection.(3)Building area dectection and contour dectection guided by the attribute of keypointThe papar desigens a method for building area detection and building contour dectection by the attribute of keypoints guided.By simulating the top-to-bottom activation process under the control of brain,utilizing the attribute of the keypoint to guide the detection mission in each levels,we can ultimately get the building contours.In the papar,we use intersections created by SLIC segmentation to induce the detection process with the attribute information together.The papar improves the classical SLIC segmentation method,merge the super pixels adjacent to the building intersection,gets and labels the minimum enclosing rectangle after the binary classfication process,and then cuts the super pixels adjacent to the non-building intersection after the three classfication,gets the initial building contour,and simplifies the initial contour.Thus it can not only improve the efficiency of computing,but also reduces the misclassification caused by a single scale.In addition,the papar introduces the vegetation segmentation model inspired by skin color detection method to improve the accuracy of the classfication further more.According to the comparison with the deep learning method for building detection using Faster R-CNN and angle classification network,this method has the similar result in accuracy,missing rate,but perfom better in terms of coverage rate.