一种鲁棒的户外建筑物图像校正方法

A Robust Outdoor Building Image Rectification Method

针对在海量场景识别中现有透视失真校正方法的鲁棒性尚不理想,不足以用于对场景复杂、干扰较多的户外建筑物图像进行准确校正的问题,提出一种改进的户外建筑物透视失真校正方法.该方法针对建筑物的形状特点,结合手机的重力感应功能对Hough变换直线检测算法进行了改进,并使用RANSAC方法剔除部分错误直线,提高了算法的鲁棒性;利用消隐点坐标对图像进行校正,恢复建筑物的仿射结构,使其与训练图像更加接近;在实验中,将校正后的图像用于图像识别,并提出一种基于互信息的融合策略,将侧视图和校正后的图像结合在一起进行识别.实验结果表明,该方法能够鲁棒地对户外建筑物图像进行透视校正,极大地提高了户外建筑物识别正确率.

For recognition of massive scenes, it can improve rate of recognition to rectify the building images with perspective distortion. The robustness of existing perspective distortion rectification methods is not enough to be used for complex outdoor building images with strong interference. In this paper, an improved perspective distortion rectification method was given. For the shape features of buildings, the gravity sensor on mobile phones was used in this method and make improvements on Hough transformation. Then RANSAC is used to get rid of some wrong lines. So the robustness of the method was improved. The horizontal and vertical vanishing points of buildings were used to restore affine structure of buildings and make them closer to the training images. In experiment, the rectified images were used to recognize images. Furthermore, a fusion strategy with mutual information was proposed to combine the result of a side view and a rectified image. The experiments showed that the proposed method can rectify outdoor building images robustly, and greatly improve correct rate of the recognition of outdoor buildings.