摘要
Light absorption and scattering in water attenuate signals, reduce contrast, and blur fine details in underwater images. When ambient light is insufficient, active illumination provides controllable incident light for the target and also generates backscatter from the water. Differences in the polarization responses of target returns and backscatter provide additional measurement constraints for physics-based image restoration. The Underwater Multi-Scene Polarization Dataset (UMPD) was developed to support research on underwater image restoration under active illumination.In the main experiments, a FLIR division-of-focal-plane (DoFP) polarization camera equipped with an IMX250MZR sensor was used to acquire 203 raw 8-bit polarization images across 21 target scenes in a water tank. A custom-built, constant-current LED source provided polarized illumination with adjustable intensity and color. The tank was covered with matte black film except at the observation window to reduce stray light and specular reflections from the tank walls. The scenes were grouped into high, low, and mixed polarization categories according to the relative polarization responses of the targets. They contain individual targets and target combinations, including manufactured objects such as a steel ruler, a paper note, coins, a rubber duck, and a plastic badge, as well as natural specimens such as a starfish, a seashell, coral, wood, and stone. Given the tank dimensions, target distances were set from 20 to 45 cm in 5 cm increments.The main experiments provide ground-truth reference images acquired in clear water and images acquired under uniform and spatially nonuniform turbidity. Skim milk was added to adjust the turbidity. Images under spatially nonuniform turbidity were acquired while the milk was still unevenly dispersed, and images under uniform turbidity were acquired after a wave maker had thoroughly mixed the milk and water. Measured turbidity under uniformly mixed conditions ranged from 6.0 to 131.0 NTU. The dataset defines low turbidity as below 30 NTU, moderate turbidity as at least 30 but below 60 NTU, and high turbidity as 60 NTU or above. In selected scenes, active illumination intensity was varied at a fixed turbidity to examine its effect on image restoration.The supplementary system comparison and temporal averaging experiments used a combined steel ruler and paper note target with mixed polarization characteristics at a distance of 30 cm. The system comparison examined the effects of white versus red illumination and of illumination intensity on polarization measurements. It also compared single-shot acquisition with the IMX250MZR-based DoFP camera and sequential acquisition with an IMX250 global-shutter CMOS camera fitted with a rotating linear polarizer. Both sensors have a 2/3-inch optical format and the same pixel size. The rotating-polarizer system acquired images sequentially at analyzer orientations of 0°, 45°, 90°, and 135° and was designed to provide high-extinction-ratio, low-noise reference measurements for comparison with the DoFP system. The temporal averaging experiment covered clear water and low, moderate, and high turbidity. Each condition includes a separately acquired single frame and a sequence of 10 consecutive frames for evaluating noise reduction through frame averaging. The supplementary data also include uncompressed raw videos of a moving steel ruler and a moving paper target, recorded at 30 frames per second at a turbidity of 40 NTU and a target distance of 30 cm. The videos preserve the original polarization samples and can be used to assess the processing latency of restoration algorithms in dynamic scenes.The dataset contains 287 images and two videos. The clear-water reference images and raw polarization measurements across turbidity conditions can be used to investigate how target polarization, scattering in water, and active illumination contribute to image degradation and to test polarimetric image formation models under uniform and spatially nonuniform turbidity. The data also support the development and evaluation of restoration approaches including physics-based iterative methods and deep learning. UMPD provides a basis for advancing polarimetric image restoration from controlled tank experiments to practical underwater vision tasks. It can support the development of vision algorithms for seafloor target recognition, underwater infrastructure inspection, and marine ecological monitoring, with the potential to improve target visibility and recognition in complex underwater environments.