基于逻辑随机共振理论的水下视觉目标检测方法研究

The attenuation of the light that travels through a water medium makes several problems of underwater images. As a result of low contrast and color distortion, images are unclear and suffering from loss of important information. Therefore, the objects in these images can hardly be differentiated from the background. This study applies logical stochastic resonance (LSR), which is typical theory to detect feeble object from heavily noisy background by the interplaying of noise and nonlinearity, to achieve the object detection task from low signal-to-noise ratio underwater images. The main framework of LSR based object detection can be described as: the heavily degraded image is placed into the LSR system as an input signal. Additional noise, as the auxiliary power to help every pixel jumping into the right state, is added to the system for separating the object from the background. Moreover, to raise the correct probability, the classical LSR structure is no longer used. A more appropriate LSR structure is constructed by choosing each module based on the inherent optical properties of turbid water. To be more adaptive to different situations, deep learning is used to optimally set the parameters of the assembled LSR system. Furthermore, LSR based object detection from turbid underwater imaging algorithm is studied, and an embedded platform is built to test the performance of the proposed method. This work aims at enhancing the capability of underwater vision detection. Meanwhile, it is also crucial for the theoretical research and engineering applications of LSR.

水下视觉探测具有重要的国防意义和经济价值。但因受水体中细微粒子影响，光学成像不清晰，尤其在浑浊水域中的目标更是被湮没于背景噪声中而难以辨认。而现有的水下图像处理方法会放大背景噪声，使目标检测更加困难。针对这一问题，本项目创新性引进逻辑随机共振思想，以像素强度为被检测量，将二维图像输入非线性势阱，配合驱动噪声增加扰动，使每个像素落入正确的状态，从而将目标与背景分离出来，以达到低信噪比目标检测的目的。为提高检测正确率，依据水体固有光学特性组合式构建逻辑随机共振系统；为提高对不同环境的适应性，引入深度学习进行逻辑随机共振系统的参数优化设置。实现可面向实际应用的基于逻辑随机共振的浑浊水域视觉目标检测算法。在此基础上，搭建嵌入式平台，进行水下遥操作机器人水下搭载实验，进一步提高算法的鲁棒性。本项目不仅能有效提高浑浊水域视觉探测水平，对于逻辑随机共振的理论研究及工程应用也具有重要意义。

随着小型化水下智能设备的发展，视觉探测，尤其是通过光学摄像头采集水下画面以完成导航、精细作业任务的需求不断增加。但由于水体对光的散射吸收，水下视觉探测距离较短。而在河流、近海等浑浊水域中，能见度往往不足30cm，在这种水域中拍摄到的画面画质模糊、色彩失真，极大的影响了视觉探测能力。传统的图像处理方法对这种非线性、高强度噪声图像处理效果不佳。针对这一问题，本项目创新性引进逻辑随机共振思想，以像素强度为被检测量，将二维图像输入非线性势阱，配合驱动噪声增加扰动，使每个像素落入正确的状态，从而将目标与背景分离出来，以达到低信噪比目标检测的目的。依据水体固有光学特性组合式构建逻辑随机共振系统，尝试使用高斯噪声、Ornstein–Uhlenbeck噪声等不同分布的噪声模拟水下噪声，提高检测正确率。同时，采用延时反馈模块来模拟散射光线造成的噪声叠加，进一步提高正确率。为提高对不同环境的适应性，我们还深入研究了水下图像的增强及复原技术，同时提出用局部清晰度指标来评价图像质量并根据其划割不同清晰度区域，针对不同清晰度采用不同的非线性系统组合。在理论研究的基础上，开展了实验室水池可控环境的实验和近海水域实验，实验证明了算法的有效性。在此基础上，搭建嵌入式平台，进行水下遥操作机器人水下搭载实验。