Underwater Depth Estimation based on Water Classification using Monocular Image

摘要

The rapid growth of computational and sensor capacities allows the development of image restoration methods that can be applied to underwater images. Due to its high degree of absorption, water becomes a major challenge for robotic perception applications. A fundamental issue for many underwater robot applications is the requirement of a depth map. One of the challenges to obtaining monocular underwater depth image is the lack of large image sets to validate the method, or even training a learning-based method. For the estimation, some methods have been proposed in the state-of-the-art either based on a physical model and on a deep learning approach. Through the analysis of the strengths and weaknesses of each kind of approach, this work aims to obtain the best depth map by classifying the input color image. For this, the water type of each image is evaluated. The results obtained in this work are promising, showing the capability of the classifier to identify the most suitable for each input image.