基于频率域、小波变换和神经网络的真彩图像增强算法研究

【作者】 熊杰；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2010， 博士

【摘要】 从真实景象中获取的真彩图像中常常既包含了刺眼的高亮区,又包含了难以看到的低亮区。为便于人眼观察和后续的机器处理,需要压缩真彩图像的动态范围,真彩图像增强。真彩图像增强技术被广泛应用于众多领域,其技术方法研究具有重要的意义与应用价值。本文针对最主流的带色彩恢复的多尺度Retinex系列方法(MSRCR)存在主要缺陷,在分析真彩图像成像的物理学和光谱学原理基础上,结合人和人眼的生理学及心理学特点,定义了一种入射——反射模型,综合运用信号处理、小波分析和神经网络等多种技术手段,建立了以频率域和小波变换为核心的真彩图像增强的方法体系,本文所提方法充分考虑到人、图像、像素之间的关系,使用效果和适应性均较为显著。主要研究进展包括：1提出了真彩图像色饱和度的调整算法。由于HSV彩色坐标系可将色彩和亮度近似分离,本文按照人眼光谱感觉特性,基于HSV坐标系,提出了一种用高斯滤波器调整真彩图像色饱和度的算法,以适应人眼对色彩偏红图像的观察。2.针对同态滤波技术不足的改进：(1)确定了同态滤波器截止频率的范围。由于入射分量中包含的细节决定了增强后图像的信息损失,用分析入射分量和反射分量方差的方法,将同态滤波方法的截止频率确定在比较合理的范围内。(2)提出一种真彩图像的多尺度同态滤波增强算法。具有单一尺度截止频率同态滤波器存在细节保持和动态范围压缩之间矛盾,用具有不同尺度截止频率的多个同态滤波器增强图像并综合,来解决这一问题。3.提出了用平滑传导函数实现图像增强的算法。Retinex方法使用中心——环绕函数对图像做卷积运算,但其缺乏灵活性,用平滑传导函数代替中心——环绕函数改进Retinex方法,有效提高了Retinex方法的灵活性,效果优于多尺度Retinex方法。4.提出了基于小波变换实现真彩图像增强算法：(1)同态分解——小波增强算法。用同态分解分离图像的入射分量和反射分量,再结合小波变换,来达到增强图像时保留细节的目的。(2)小波——能量增强法。针对传统入射——反射模型使用上的限制,采用静态小波变换定义一个新的入射—反射模型,给出了图像增强是一个图像能量衰减和再放大过程的新思路,使用时对需要增强的真彩图像基本没有条件限制,其适应性和效果明显优于MSRCR。(3)设计实现了能同时对真彩图像增强和降噪的算法。分析了真彩图像噪声在HSV坐标系中的分布情况,在使用小波——能量法增强图像基础上,用贝叶斯软阈值滤除被分离在反射分量中的噪声,以达到增强图像时抑制噪声的目的。5.基于神经网络的真彩图像增强方面：(1)提出小波——PCNN(脉冲耦合神经网络)增强算法。根据韦伯定理,采用PCNN确定图像调整的指数放大系数,图像中具有不同亮度的部分用不同的放大系数,在同一幅图像中做到衰减高亮区和增强低亮区,其方法适应性和效果优于小波——能量法。(2)提出递归神经网络恢复真彩图像色彩的算法。用递归神经网络的记忆功能记忆图像色彩,建立了用于色彩校正的递归神经网络的权值矩阵和方法,校正在RGB坐标系中使用同态滤波增强图像时产生的色彩偏离。更多还原

【Abstract】 Many real color images, which are photographed from real scenes, possessing high dynamic range include both dark shadows and bright light sources that are difficult to be perceived. The dynamic range of these images should be compressed (that is real color image enhancement) in order that these images are perceived by humans or rendered more suitable for machine analysis. Because real color image enhancement is widely applied in many domain, the research of real color image enhancement is the important worth of application. Because the color sustainment of images enhanced and the application of illumination-reflection model are confined by MSRCR and real color images do not uaually satisfy the demand of MSRCR, the default of "halos", images covered by mist and the actual scenic details and/or colors obscured are common in real color images enhanced. According to the physical and spectroscopic theory of real color images and the physiological theory of human and human eyes and the phychological theory of human vision, we create a new llumination-reflection model and put forward a series of algorithms based on digital signal processing, wavelet transform, neural networks and etc. The flexibility and validity of these algorithms are good.The paper includes five aspects.1. The algorithm that the saturation of real color image is adjusted is put forward. Because HSV color system can approximately separate the color and the value of real color images, according to the spectral sensitivity of most people vision, the saturation of real color image with dark red in HSV can be adjusted by a Gaussian filter to accommodate the observation of human eyes.2. To improve homomorphic filter:(1) The cut-off frequency range of homomorphic filter is confirmed. Because the illumination determine the information of image enhanced, we can analyse the variation of the illumination and the reflection in order that the cut-off frequency is decided in a reasonable scope. (2) A algorithm of real color image enhanced by multi-scale homomorphic filters in two channels is put forward. The method of multi-scale homomorphic filters is put forward in order to eliminate the contradiction between the images detail sustained and the images high dynamic range compressed. 3. A smoothing conduction function method is put forward to enhance images. Retinex is improved by a smoothing conduction function in order that the flexibility of Retinex is increased. The method is better than Retinex.4. Real color image enhanced based on wavelet transform:(1) A homomorphic decomposition-wavelet transform algorithm is put forward. The illumination and the reflection are separated by homomorphic decomposition. The detail is sustained by wavelet transform. (2) A wavelet-energy enhancement is put forwand. A new illumination-reflection model is described by stationary wavelet transform in order to eliminate the limitation of old illumination-reflection model. A new idea is that image enhancement is a process that the imagery energy is reduced and then amplified. Because real color images enhanced by the method are not restricted, the method is much better than MSRCR in flexibility and effect. (3) Real color image is enhanced and denoised by stationary wavelet transform at one time. We analyse the noise of real color image in HSV system. The noise in the reflection is eliminated by bayes-soft-threshold when real color image is enhanced by wavelet-energy.5. Real color image enhanced based on neural networks:(1) A wavelet-PCNN (pulse coupled neural networks) is put forward. According to Weber theory, the exponents which adjust the imagery value are decided by PCNN. Different parts possessing different luminance in image are amplified by different exponent. The method is called Gamma adjustment in different levels. High bright parts are reduced and shade parts are amplified at the same time in one image. The method is better than the wavelet-energy enhancement. (2) RNNs (recurrent neural networks) revise the color of real color images enhanced. RNNs (recurrent neural networks) possess the recall ability. Based on the analysis of RNN’s stability and convergence, weight matrix of recurrent neural networks is properly confirmed to revise the color real color images enhanced in the paper.更多还原