Road Extraction On Fog And Haze (Low Quality) Aerial Image

In recent years,aerial photography technology has become one of the principle measures to obtain ground feature information,especially road information.However,because of the increasing frequency of hazy weather,the picture of the targeting road taken via aerial technology could be difficult to display,which will seriously affect the related recognition and extraction in the late work.This paper focuses on research about the road extraction algorithm,trying to address the problems like blurring,low-contrast,dimming and etc.of the aerial image leading by hazy weather.This paper contains two parts,preprocessing of aerial image and road information extraction.The preprocessing is the foundation of road extraction,and road extraction is the purpose of the preprocessing.Both of them are complementary and form an organic whole.The aerial image preprocessing includes two steps: clearness and enhancement.And road extraction part contains two methods,one is common method,the other is aim at ridge feature.The following sections are involved:1 On the foundation of relevant theories regarding defogging aerial image,especially the research and analysis of Retinex algorithm,this paper proposes a new method of Retinex algorithm based on the change of the scene depth scale.To be specific,Retinex algorithm and image transmission diagram are combined,and different scales of Gaussian filter function is applied for different regions according to scene depth information.Through enormous experiment analyses and comparison,adopting average value,standard deviation and average gradient to make object evaluation,which shows that the proposed method can effectively improve the quality of the image.The processed image is more clear and natural in color,and with better effect of defogging.Above all these,this method could avoid the certain defects of Retinex algorithm,laying a good foundation for next stage,target extraction.2 In order to enhance the fog and haze(Low Quality)aerial image and clear the details information,the fractional order differential operator is applied.After the analysis of the principle and characteristics of Tiansi operator,the improved operator coefficient of each pixel in the template is calculated by the length of the distance testing point.At the same time,the sum of every pixel in the template is not zero,and the size is determined by the neighborhood,distance to center and fractional order.The improved operator can enhance the image detail information,ensure little changes on the brightness,and restrain the noise.The objective criteria adopted to evaluate the enhancement effect are image information entropy and image histogram.By the contrast to other method,the experiment shows that the proposed temper can achieve better results.3 After the clearness and enhancement for fog and haze aerial image,a road segmentation algorithm based on K-means clustering and MST is proposed.K-means algorithm is adopted to achieve the initial clustering,and then the MST algorithm is applied to extract roads information.Due to the characteristic of aerial image,i.e.more noise,unbalanced illumination and tiny margin,this proposed algorithm combines the graph theory and K-means clustering to reduce the over segmentation because of MST.Then,the related post-processing methods are selected to remove the flaws of the image and extract accurate road information from the aerial images.After comparing with the traditional algorithm,the proposed method can obtain better extraction.4 In order to extract road information with ridge features from aerial images,an improved ridge edge detection algorithm is proposed.Compared with other common edge algorithms,this algorithm is not based on single detection,but based on a 3-4 points to detect short lines,which can ensure the targeting edge belonging to ridge edge and improve the detection accuracy.Compared with other conventional edge detection algorithm,the results show that this proposed algorithm reaches good extraction effect for road information with ridge features,and a realistic balance between noise immunity and detection accuracy.