基于主动推理的水下自主航行器视觉认知计算关键技术研究

21世纪是人类开发和利用海洋的新世纪，水下自主航行器海洋探测技术已成为各国争相投入的极具挑战性的前沿技术。本项目针对目前水下自主航行器对动态目标缺乏有效特征分析模型问题，提出结合时空注意力特征网络和深度变分贝叶斯自编码器的视觉小样本数据增强模型，为水下自主航行器视觉认知计算方法提供必要的基础支撑；针对目前水下复杂视觉环境高效智能决策推理能力不足难题，提出基于深度主动推理的水下图像认知计算模型及加速硬件架构设计方法，增强水下自主航行器视觉探测分析与决策能力；基于上述研究成果，提出一套支持复杂水下环境、低功耗、高处理能力的基于主动推理的视觉认知理论与方法，构建水下自主航行器视觉认知计算原型系统并进行试验验证。力争在视觉探测认知领域取得高水平的研究成果，推动高层次人工智能技术的发展与应用，具有重要的研究价值与现实意义。

The 21st century is the new era for ocean exploitation, and the marine detection technology of autonomous underwater vehicles has become a challenging technology and an area of competing investment by countries. For the lack of effective feature analysis for dynamic targets of autonomous underwater vehicles, this project puts forward a network enhancement model combined with an attention feature network based on space-time image data and the deep variational bayesian auto-encoder network. In order to improve the inadequate ability of intelligent decision-making in the complex visual underwater environment, this project will propose a model of underwater image cognitive calculation by the deep active inference, which make an enhancement on the ability of visual detection and decision-making. Based on the research above, we get a new design method for visual cognitive supporting underwater environment and low power consumption with high performance, and this work will establish a computing prototype system for visual cognitive of autonomous underwater vehicles with a validation. We will strive to obtain high-level research results in the field of visual cognition. It is of great research value and practical significance to promote the development and application of artificial intelligence technology.