New shadow detection and removal approach to improve neural stereo correspondence of dense urban VHR remote sensing images

Abstract

Shadows cause problems in many remote sensing applications like images segmentation, objects extraction and stereo vision. This paper presents a new and an automatic approach to detect and remove shadows from pair of dense urban very high resolution (VHR) remote sensing images. The main contribution of this paper is twofold. First, a proposed approach is efficient to restore objects hidden by shadows, second, it improves a stereo matching process. We have chosen to operate on Ikonos pairs as an example of urban remote sensing images, for that, shadow detection is achieved using a new technique of property based method, operating directly in red. green and blue colour space (RGB). Shadow removal proposed technique aims to produce a needed amount of light to the shadow regions by multiplying the shadow regions by constants, after that, the shadow edge correction is applied to reduce the errors due to diffusion in the shadow boundary. Once pair of shadow free images is recovered, we apply a stereo matching process using a Hopfield neural technique in order to find homologous regions. Our results from different urban pairs show the effectiveness, the simplicity and the fastness of the proposed approach to reveal details hidden by shadows and to obtain a high stereo matching rate.