Line Detection from Photo Images and Image Retrieval using Line Information

Abstract

In almost all photo images, one can see various straight lines such as horizontal lines, the outines of tree trunks, buildings, door frames and so on. This line information can be used to characterize these images as well as the objects in the images. This paper proposes a new line detection algorithm based on the Hough transform. We improve the efficiency fo parameter estimation by using edge direction information and reduce the computation time with an edge thinning method. Experimental results on 1010 PhotoDisc images show that 40 percent of detected lines can be utilized for image retrieval. We study the image similarity retrieval which utilizes the detected lines as features of images. In order to retrieve images of various aspect ratio, we use a new method that normalizes the offset of the lines. We implement a simple image retrieval prototype and show that the normaization method works effectively in retrieving the images containing lines similar to the reference specified by a user.