This paper presents an algorithm detecting automatically a wide variety of cracks on monochrome images of concrete floors. It is part of a vision system, generating a crack map to support the condition monitoring for buildings. The suggested method uses successively radiometric, geometric and contextual information. An automatic supervised adaptive intensity-threshold method handles radiometry information. A threshold-based method was chosen because it can separate even thin cracks from the background. In order to cope with different image cluster intensities, first a clustering algorithm separates regions of different intensities, and second the threshold is adaptive at the cluster level; being a continuous function of the cluster intensity. At some points, function values were learned via supervised classification. Then, we performed an interpolation between these points in order to get a threshold whatever the cluster intensity (continuity). At this step, we have a thresholded image. However, due to texture, non-crack dark defects and dirtiness we have false positives. In order to overcome this problem, connected pixels are grouped into regions, and after discarding small regions, a size-dependent geometrical shape filter is suggested. The shape of a crack region depends on its area, and this relation was empirically learned. A region is retained if its shape features are above some area depending thresholds. However, mainly due to texture, we have still false positives because some non-crack entities have radiometry and geometry of cracks. Fortunately, they are often small isolated regions and are discarded via an isolation filter. Tests performed on many images show very encouraging results.

This paper presents a 3D vision sensor and its algorithms aiming at automatically detect a large variety of defects in the context of industrial surface inspection of free-form metallic pieces of cars. Photometric stereo (surface normal vectors) and stereo vision (dense 3D point cloud) are combined in order to respectively detect small and large defects. Free-form surfaces introduce natural edges which cannot be discriminated from our defects. In order to handle this problem, a background subtraction via measurement simulation (point cloud and normal vectors) from the CAD model of the object is suggested. This model-based pre-processing consists in subtracting real and simulated data in order to build two complementary "difference" images, one from photometric stereo and one from stereo vision, highlighting respectively small and large defects. These images are processed in parallel by two algorithms, respectively optimized to detect small and large defects and whose results are merged. These algorithms use geometrical information via image segmentation and geometrical filtering in a supervised classification scheme of regions.

A method is proposed for building and road detection on very high spatial resolution multispectral aerial image of dense urban areas. First, objects are extracted with a segmentation algorithm in order to use both spectral and spatial information. Second, a spectral-spatial object-level pattern is formed, and then classification is performed using a 3-class SVM classifier, followed by a post-processing using contextual information to handle conflicts. However, in the particular case where many building roofs are grey like the roads and have similar geometry, classification accuracy is inevitably limited. In order to overcome this limitation, different classifiers are combined and different patterns used, improving the accuracy of 10%.

A method is proposed for building and road detection on VHR multispectral aerial images of dense urban areas. Spatial and spectral features of segmented areas are classified using a 3-class SVM integrating some a priori and contextual information to handle unclassified patterns and conflicts. Geometrical object features and additional information improve the classification accuracy in the difficult case where many building roofs are grey like the roads and have similar geometry. Also, road network regularization is suggested to improve the classification accuracy.

Marked watershed transform can be seen as a classification in which connected pixels are grouped into components included into the marks catchment basins.The weakened classifier assembly paradigm has shown its ability to give better results than its best member, while generalization and robustness to the noise present in the dataset is increased. We promote in this paper the use of the weakened watershed assembly for remote sensed image segmentation followed by a consensus (vote) of the segmentation results. This approach allows to, but is not restricted to, introduce previously existing borders (e.g. for the map update) in order to constraint the segmentation. We show how the method parameters influence the resulting segmentation and what are the choices the practitioner can make with respect to his problem. A validation of the obtained segmentation is done by comparing with a manual segmentation of the image.

Homographies are widely used in tasks like camera calibration, tracking, mosaicing or motion estimation and numerous linear and non linear methods for homography estimation have been proposed in the case of classical cameras. Recently, some works have also proved the validity of homography for catadioptric cameras but only a linear estimator has been proposed. In order to improve the estimation based on correspondence features, we suggest in this article some non linear estimators for catadioptric sensors. Catadioptric camera motion estimation from a sequence of a planar scene is the proposed application for the evaluation and the comparison of these estimation methods. Experimental results with simulated and real sequences show that non linear methods are more accurate.