水下成像中的光谱特征提取研究

Study on spectral feature extraction in underwater imaging

针对当前基于最小二乘的水下成像中的光谱特征提取方法,存在的水下成像的复原质量较差,图像的去噪效果,和光谱特征提取效果不是十分理想等问题。文章提出利用曲线拟合和下采样法,实现水下成像中的光谱特征提取。通过水下成像采集的简易装置,对水下图像进行采集;通过合理的假设,和光学理论的公式推导,对背景光需要的水体光学方面的参数进行计算,根据散射系数和波长之间的关系,对R、G、B通道的透射率进行计算,同时根据导向滤波对投射率图进行精细化,利用逆求解的成像模型,对水下图像进行复原,以达到增强水下图像的复原质量的目的;利用改进软硬阈值的折中法,针对水下图像的特性,完成图像的去噪,在图像复原的基础上实现图像的双重去噪,提升去噪效果;根据曲线拟合和下采样法,对水下图像光谱的拟合结果,实施数据冗余的清除,以及光谱特征的提取操作,并且在实现中,只在少数若干波长处,对图像光谱的流量进行高效采样,不仅可以提升光谱特征提取效果,还能够提高光谱特征的提取速度。实验证明,所提方法无论是图像复原方面,还是图像去噪方面,或者图像光谱特征提取方面,均优于当前方法,以此说明了该方法可行性较好。

Aiming at the current methods of spectral feature extraction in underwater images based on least squares,the underwater imaging quality is poor,the image de-noising effect and the spectral feature extraction are not ideal. The article proposes the use of curve fitting and down sampling method to realize the spectral feature extraction in underwater imaging. Based on the reasonable assumptions and formulas derived from optical theory,the optical parameters of the water body for the background light are calculated. Based on the relationship between the scattering coefficient and the wavelength,the transmittances of R,G,B channels are calculated,and the projection rate map is refined according to the guide filtering,and the underwater image is reconstructed by using the inverse solution imaging model so as to enhance the restoration quality of the underwater image. According to the characteristics of underwater image,the image de-noising is completed,and the double de-noising of image is realized on the basis of image restoration to enhance the de-noising effect. According to the curve fitting and down sampling method,fit the results of underwater image spectra to remove data redundancy and feature extraction. Actually,effective sampling the flow of image spectra at only a few wavelengths can not only enhance spectral feature extraction effect,but also improve the spectral feature extraction speed. Experiments show that the proposed method is better than the current method in terms of image restoration,image de-noising and image spectral feature extraction,which proves the feasibility.