Challenges of Close-Range Underwater Optical Mapping Ricard Prados, Rafael Garcia Computer Vision and Robotics Group University of Girona, Girona, 17001 Spain Email: {rprados, rafa}@eia.udg.edu Javier Escart´ ın ´ Equipe de G´ eosciences Marines CNRS/IPGP UMR7154 75238 Paris, France Email: escartin@ipgp.fr L´ aszl´ o Neumann Computer Vision and Robotics Group University of Girona, Girona, 17001 Spain ICREA, Barcelona, 08010 Spain Email: lneumann@eia.udg.edu Abstract—Underwater optical mapping often involves the use of image mosaicing techniques. High quality mosaicing requires the application of blending methods to achieve continuous and artifact-free mosaics. Image blending has a dilated history of over three decades in the terrestrial and aerial fields. Unfortunately, the nature of the underwater medium adds additional difficulties to the mosaicing and blending tasks. In this paper a survey of the blending methods is given, focusing the attention on its applicability to underwater mosaicing. Image acquisition is performed by Autonomous Underwater Vehicles (AUVs) or Remotely Operated Vehicles (ROVs) in the deep ocean, a medium with aggressive light absorption and disrupting scattering effects that requires of the use of artificial lighting. A comprehensive comparison of the basic features and limitations of some of the most important existing blending techniques is presented. The goal is the generation of seamless and visually pleasant large area photo-mosaics of the seafloor, free from double contouring, ghosting and other disturbing and common blending artifacts. I. I NTRODUCTION High resolution optical imaging provides scientists (e.g., archeologists, geologists, biologists, among others), with ac- curate and rich visual information of the ocean seafloor. Object measurement, monitoring and terrain prospection are among the tasks which can be performed based on underwater photographs. Despite increased resolution of nowadays electronic still cameras, underwater regions of interest are usually too ex- tensive to be covered with the required level of detail on a single shot. Furthermore, due to light attenuation, seafloor images should be acquired at short range in order to minimize its impact. Consequently, several images should be merged in order to obtain a picture that covers the whole surveyed area. Moreover, this approach needs to deal with additional difficulties such as non-uniform illumination, light scattering and moving objects (see Fig. 1). All these effects cause strong visual artifacts when several images are stitched together to form a photo-mosaic. Image blending techniques are widely used in indoor and outdoor contexts in order to generate seamless panoramas. Most of the conventional algorithms work on the basis of a still camera with rotation and negligible translation between images. In the underwater medium the phenomena described above constrain the acquisition range to a few meters over the seafloor, commonly using artificial light sources, which leads to illumination artifacts but also to geometrical problems such as parallax even when the 3D relief of the seabed is small. Despite the existence of 3D reconstruction methods intended to underwater imagery [1], [2], [3], which allow avoiding parallax issues, those ones cannot be applied when the overlap between images is small. That is the case of the datasets used in this work, acquired by autonomous or teleoperated vehicles equipped with still cameras that are synchronized with stroboscopic flashes as lighting sources. The stroboscopic light requires about 10 seconds to recharge and shot, due to its high level of power, restricting this time interval the maximum acquisition rate. Quality blending is important inasmuch as a scientist can significantly improve its visual analysis and data interpretation when working with a continuous and uniform high resolution photo-mosaic instead of a simple sequence of stitched images, i.e., a non-blended mosaic, or the individual images. Besides the image quality of the mosaic, the size of the data to process, concerning picture size and number of pic- tures, is another challenge by itself. The large photo-mosaic dimensions affects not only its processing time but also the computer memory management during its processing, requir- ing a tailored strategy for the agile processing and visualization of the results. The aim of this work is to review the current image blending techniques aimed at mosaicing. We point out their capabilities and weaknesses when applied in the underwater medium. The target of the study is to develop an adequate processing pipeline to deal with huge and challenging underwater image sequences in order to build high quality large scale photo- mosaics of the seafloor. II. STATE- OF- THE-ART BLENDING APPROACHES A. Literature Review The basic principles of image blending where established four decades ago [4] and include two main concepts which drive to two algorithm groups [5]: transition smoothing and optimal seam finding. On the one hand, transition smoothing methods [6], [7] fade the images along a common overlapping region in order to minimize the visibility of the seam. On the other hand, optimal seam finding algorithms [8], [9] focus on computing the joining boundary which reduces in 978-1-61284-4577-0088-0/11/$26.00 ©2011 IEEE