In this letter, we propose a novel automatic algorithm for road extraction from remote sensing images. The algorithm includes low- and high-level processing. In the low-level processing, we determine a normalized second derivative map of road profiles of a generalized bar shape, which is width invariant and contrast proportional, and accordingly obtain initial road center pixels. In the high-level processing, using the map and initial center pixels, we initially determine road segments. The segments are then locally refined using their orientation randomness and length-to-width ratio and further refined via global graph-cut optimization. A final road network is thereby extracted in a robust manner. Experimental results demonstrate that the proposed algorithm provides noticeably more robust and higher road extraction performance in various images compared with the existing algorithms.

A digital elevation model (DEM) contains the elevation of terrain, which is widely used in the geographic information system (GIS). However, a DEM may not be accurate due to the measurement inaccuracy and the change of topography with time. In this paper, we propose a feature based algorithm for detecting DEM elevation errors using a camera image. In the algorithm, we first extract ridges in a depth map obtained from a DEM. We then extract ridges in the corresponding camera image, and determine the elevation error in the DEM, if any, by comparing the corresponding points of ridges between the depth map and camera image. To robustly extract a ridge in the camera image, we adopt a mask that is obtained using the ridge information in the depth map, and apply dynamic programming within the mask. Experimental results show that the proposed algorithm can reliably determine elevation errors in the DEM.

In this paper, we propose a novel pixel-level multi-sensor image fusion algorithm with simultaneous contrast enhancement. In order to accomplish both image fusion and contrast enhancement simultaneously, we suggest a modified framework of the subband-decomposed multiscale retinex (SDMSR), our previous contrast enhancement algorithm. This framework is based on a fusion strategy that reflects the multiscale characteristics of the SDMSR well. We first apply two complementary intensity transfer functions to source images in order to effectively utilize hidden information in both shadows and highlights in the fusion process. We then decompose retinex outputs into nearly nonoverlapping spectral subbands. The decomposed retinex outputs are then fused subband-by-subband, by using global weighting as well as local weighting to overcome the limitations of the pixel-based fusion approach. After the fusion process, we apply a space-varying subband gain to each fused subband-decomposed retinex output according to the subband characteristic so that the contrast of the fused image can be effectively enhanced. In addition, in order to effectively manage artifacts and noise, we make the degree of enhancement of fused details adjustable by improving a detail adjustment function. From experiments with various multi-sensor image pairs, the results clearly demonstrate that even if source images have poor contrast, the proposed algorithm makes it possible to generate a fused image with highly enhanced contrast while preserving visually salient information contained in the source images.

In color image processing, an image is usually processed in a color space consisting of luminance and chrominance components. The processed result is then converted to the RGB color space for display. However, a straightforward color space conversion can cause color distortion due to gamut difference between the two color spaces. In order to alleviate this gamut problem, we propose two gamut-adaptive correction schemes. For the chrominance component processing, we propose the gamut-adaptive clipping scheme to effectively solve the out-of-gamut problem. For the luminance component processing of a color image, we propose the gamut-adaptive scaling scheme to alleviate both the out-of-gamut and under-saturation problems. The proposed methods can be applied for image processing in any color space since the proposed correction is performed in the RGB color space. Experimental results show that the proposed methods can preserve the luminance and hue values and provide rich colors by adaptively controlling the saturation value according to the luminance component value.

This paper presents a novel multi-sensor image fusion algorithm to generate a contrast-enhanced fused image. The algorithm is based on a framework of the subband decomposed multiscale retinex (SD-MSR) that is one of the contrast enhancement methods. It also uses two complementary intensity transfer functions to enhance the details in both shadows and highlights. For effective contrast enhancement, we suggest a novel fusion strategy which reflects the characteristics of SD-MSR well. Thereby, even if input images have poor contrast, the proposed algorithm is able to generate a fused image with highly enhanced contrast. Experimental results demonstrate that the algorithm performs well for generating high contrast fused images.