子計畫四：水下人工智慧影像辨識與重建(2/2)

Autonomous Underwater Vehicle (AUV) plays a pivotal role in national security, resource exploration, species recognition, submarines pipeline repair, life rescue etc., and its related applications are increasing. Given that artificial intelligence is rapidly blooming and image sensor has become one of the most important sensors of AUV, the development of deep learning-based underwater image processing algorithms has become an important trend. To break the bottlenecks of current technologies and to expand the application scope of AUVs, we aim to develop groundbreaking algorithms for underwater image recognition, restoration, and 3D reconstruction. In the year 2018, we have developed an efficient underwater object detection network, and proposed a color restoration network to solve the problem of underwater light absorption and scattering effect. In the following year 2019, we proposed an over-land image dehazing algorithm which will be adopted as the background architecture of underwater image dehaze method. Furthermore, the HDR-based SLAM has been developed for camera pose estimation. According to the experimental results, the proposed methods consist of the following four advantages: (i) The proposed underwater object detection network can work in low-power embedded system with high accuracy and low computing cost. (ii) The color restoration network can generate natural colors and completely restores the attenuated colors of underwater images. (iii) The over-land dehazing network can greatly maintain the texture and color information of the dehazed images. (iv) The HDR-based SLAM method has good performance under challenging low-light environment, which helps expand the applicability of 3D reconstruction system. By integrating and optimizing the developed technologies in these two years, the AI system for underwater image recognition and reconstruction will be built to improve the abilities of AUVs in various underwater environments.