【作者】 孙飞飞；

【导师】 王国宇；

【作者基本信息】 中国海洋大学 ， 计算机应用技术， 2011， 博士

【摘要】 数字图像处理技术的发展在很多领域已经取得了显著的成果,但是针对水下图像处理的研究却鲜有成效,这是由于水下成像的环境比陆上复杂得多。虽然海洋光学的研究历来已久,但因为光在水介质中传输衰减这一物理限制而长期受到冷落。近十几年来关于水下视觉技术的研究驱动主要来自迅速增长的对海洋探测开发的手段需求。高分辨率的视觉模式本身具备传统声纳所无法替代的优势,而计算机视觉技术的进展也使得水下视觉在海洋探测方面得到越来越多的应用。然而,改善水下成像质量的图像处理方法和技术目前还远远不能满足和应用需求。除了水介质的吸收以外,悬浮粒子对光的散射效应是限制水下观测距离的关键因素。由于水下没有环境光源,成像系统必须依赖主动照明方式。目标对照明光线的反射在向传感器的传输过程中被散射而导致成像模糊称为前向散射,根据海洋光学小角度散射理论,可以用点扩散函数(PSF)或光学传递函数(OTF)来描述前向散射效应。而照明光线在到达目标之前同样会受到水体的散射而形成后向散射光被传感器接收。后向散射效应在图像中造成一种“雾化”背景,导致图像对比度下降。前向散射和后向散射的同时作用导致水下图像严重降质,这是限制水下观测距离的主要原因。虽然提高照明功率可以增加目标反射光强度,但是后向散射光的强度也同时增加,并不能提高图像对比度。因此,图像处理技术是提高水下成像观测距离的必要手段。本文从图像增强和图像复原两个方面对改善水下图像质量的方法进行了系统研究。论文的主要工作包括以下几方面：(1)提出了一种基于散射分层传输原理的PSF模型和后向散射统计描述方法。以往的水下图像复原方法中,PSF的模型是根据水下小角度散射理论以及经验公式得到,或者通过基于传感器的图像特征提取得到数值化描述。但是这些方法都没有给出关于后向散射这一导致图像降质的重要因素的定量描述。在理论分析和实验结果的基础上,本文从后向散射与前向散射的物理机制出发,建立了一种新的散射分层传输描述方法,通过将目标与接收器之间的水体分层为各个独立的散射单元,并将各单元的散射传输视为各个线性子系统的输出,推导出了一种简化的PSF模型,在此基础上给出了后向散射噪声的统计描述。这一模型的建立,为本项研究中水下图像增强和复原的实现提供了原理框架。这是本文的主要创新点。(2)研究了抑制后向散射噪声背景的图像增强方法。传统的图像去噪方法一般把噪声视为不相干的白噪声。本文的理论分析和水池实验结果表明,后向散射信号是由直流分量和随机起伏噪声构成的,排除直流分量的影响,随机起伏噪声呈现强烈的相干性,基本特性是后向散射噪声能量集中在低频部分。其原因正如我们所提出的散射分层传输模型所描述的,后向散射是多层散射信号经过不同长度的水体所对应的低通滤波后的叠加。因此,直接应用现有的去噪方法不能解决水下图像的后向散射噪声抑制问题。在研究把现有的小波去噪方法应用于水下图像时,可以发现虽然高频噪声分量得到抑制,但是与目标信号能量重叠的占主要地位的噪声低频分量仍然无法消除。我们的研究结论是：水下图像后向散射噪声抑制的原则是高通滤波而非传统去噪方法所依据的低通原则。根据这一结论,本文提出了小波去噪结合高通滤波以及利用传统图像增强滤波抑制后向散射的方法。(3)提出了一种基于估计的水下图像复原方法。图像复原理论为解决水下图像降质问题提供了原理性框架。有关水下图像复原的研究一直集中在PSF的描述上,模型都是建立在海洋光学理论和小角度前向散射理论基础上的半经验公式。但现有模型的一个共同特点就是均依赖于水下光学参数的先验知识,这些先验知识是需要标准的海洋光学测量加上理论计算来完成的,因此更适用于有代表性的特定海域。而对于在近海作业中的实时观测方面的应用,动态环境的变化使得先验知识不可用,而现场测量经常是一个高代价的技术难题。为此,我们提出一种基于估计的图像复原方法,可以在实时、动态的条件下,仅通过对后向散射信号的分析对目标图像进行近似复原。这一方法的实现源于我们所提出的基于散射分层传输模型的后向散射描述,只要通过现场测量后向散射背景,即可以获得复原所需的系统参数而无需关于水下固有光学参数的先验知识。(4)本文提出一种基于后向散射噪声物理模型的水下低信噪比条件下的目标检测方法,并通过实验验证该方法的有效性。与传统的水下弱目标检测方法相比较,本文采用的方法无需知道水下目标的尺寸和灰度级以及水下图像的对比度。因此,该方法更加灵活且具有更强的应用价值。更多还原

【Abstract】 Digital image processing technology has achieved remarkable results in many areas, but the study of underwater image processing is rarely effective, which is due to the environment of underwater imaging is much more complex than on land^[1]. Although the ocean optics research has always been studied for a long time, it was also ignored in a long term because of the physical limitations that the attenuation when light transmited in the water. Over the last decade, the rapid growth of demand for ocean exploration and development is the mainly driving force of underwater vision. The high-resolution visual model has the advantages which can not be replaced with the traditional sonar, and the development of computer vision makes the ocean underwater exploration to get more and more applications. However, the method to improve the quality of underwater image processing technology can still is far from meeting the requirements of applications.The key factor except absorption of water medium is scattering effect of suspended particles. Because of the absence of light source, imaging system must rely on the active lighting pattern. Reflection light of illumination is scattered when it transmits from object to the sensor, which is called forward scattering and it causes the image blur. According to ocean optics’small-angle scattering theory, we can use point spread function （PSF） or optical transfer function （OTF） to describe the forward scattering effect. Before reaching the target, the illumination rays are also scattered by the same water body to form backward scattering, and then the back-scattering is received by the sensor. Backscattering effect occur a "fog" background in the underwater image, which results in the contrast of the image decreased. The combined action of backscattering and forward scattering can lead the underwater image degraded seriously, which is the main reason to limit the distance of underwater observation. While increasing the illumination power can increase the intensity of object reflected light, but the intensity of backscattering is increased too, so the image contrast can not be improved. Therefore, the image processing technology is the necessary means to improve underwater imaging system.In this paper, the method of improving underwater image is discussed from the two aspects, which are image enhancement and image restoration. And the main study of the thesis are listed as follows:1. We proposed a simplified multi-layer transfer model to formulate both the PSF and the statistics of backscattering. In the previous methods of underwater image restoration, the model of PSF was acquired according to the theory of underwater small-angle scattering and experience formula. And the value description was acquired by the image feature extraction of sensor. However these methods did not describe the backscattering quantitatively, which is the important factor to cause image quality degraded. Based on the theoretical analysis and experimental results, we established a new description method of scattering layer-transmission proceeding from the physical mechanism of backscattering and forward scattering. With separating the water body between object and receiver into many independent scattering units and regarding the each scattering unit as an output of each linear subsystem, a simplified model of PSF was derived and the statistical description of backscattering noise was also gained on the basis. The model provided a framework of principles for underwater image enhancement and restoration in this study. So it is the main innovation of this article.2. We studied the enhancement methods for restraining the backscattering noise in the paper. The traditional image denoising methods focused on the irrelevant white noise usually. The theoretical analysis and experimental results of this paper showed that the backscattering signal was composed of the direct component and random fluctuation noise. Excluding the impact of the direct component, the random fluctuation noise showed a strong coherence, and its basic characteristics was that the energy of backscattering noise was concentrated in low frequency. The reason, as we described in the scattering multi-layer transfer model we proposed, was that backscattering was a superposition of the multi-scattering signals through the low-pass filter corresponding to the different length water. Therefore, it did not suppressing the backscattering noise of underwater images by using the existing denoising methods directly. In the study of the existing wavelet denoising method that was applied to underwater images, although the high-frequency noise components can be inhibited, the low-frequency of noise which is the main noise and overlaps with the object signal still can not be suppressed. Our conclusion is that the principle of underwater image backscattering-noise suppression is high-pass filter instead of the low-pass filter principle in the traditional denoising methods. Based on this conclusion, we presented a method to suppress the backscattering noise by using the method of combining wavelet denoising with high-pass filtering and traditional image enhancement filtering.3. We proposed an underwater image restoration method based on estimation. The theory of image restoration provided a schematic framework in order to solve the problem of the degraded underwater image. The researches about image restoration has always focused on the description of PSF, the models were all built on the basis of the ocean optics theory and small-angle forward scattering theory. The common feature of the existing models is that they all depend on the prior knowledge of underwater optics parameters, and the prior knowledge need to be acquired with standard ocean optical measurements and theoretical calculations, so the existing models are more suitable for the particular representative marine area. For the application of the real-time observation in offshore operations, the changes of dynamic environments make the prior knowledge unavailable, and on-site measurement is often a costly technical problem. To this end, we proposed a method of image restoration based on estimation which can restore the target image approximatively under the real-time and dynamic conditions only by analyzing the backscattering signals. The implementation of this method dued to the backscattering description based on the scattering multi-layer transfer model. Only with the in-site measurement of backscattering background, we can get the system parameters needed by restoration without any prior knowledge about underwater inherent optical parameters.4. We proposed a method of the target detection with the low SNR, and the method is based on the power spectrum of underwater backscattering noise. By using this method, we can obtained a better result. And the method needs no contrast of image or gray scale and size of target, so it is better than the traditional methods.更多还原