图像与视频自适应去雾关键算法研究

发布时间：2018-03-01 11:35

本文关键词： 图像复原多尺度分析图像去噪视频去雾　出处：《华侨大学》2017年硕士论文　论文类型：学位论文

【摘要】：雾天环境下,大气中随机介质,如悬浮颗粒、雾、霾等对户外场景的辐射影响导致能见度严重不足,导致相机设备拍摄的画面存在对比度低、色彩暗淡等不同程度的退化问题,影响人们的视觉观感和有用信息的自动提取。因而,复原清晰这些降质图像具有十分重要的现实意义和应用价值。目前去雾方法在自适应性上还不足,难以处理远近不同场景以及噪声干扰情况下的雾天降质图像和视频。本文分析了雾天成像的物理退化模型,详细探索了图像去雾用到的关键技术,设计并实现了图像和视频去雾方法。主要研究和贡献如下:针对传统基于暗通道去雾中离视点较远区域出现的色彩失真,较近区域出现晕影等问题,提出基于场景深度分割和开暗通道的去雾方法。方案如下:首先检测输入图像场景的深度信息,划分整个图像为近景和远景两部分;采用开暗通道模型估测大气光幕,修正远景区域的透射率值;融合近景和远景优化透射率图,最终实现有雾图像的清晰化。实验结果表明,针对不同场景的有雾图像,该方法能切实有效克服暗通道先验方法的不足,视觉效果得到了显著提升,自适应强。观察到雾趋近分布于图像的低频,提出结合多尺度小波分析的去雾方法。该方法先将图像从空域变换到频域,对低频区域去雾;鉴于雾对高频信息的削弱,亦需对高频细节等同复原。方案如下:首先使用小波将图像分解到小波域,得到图像的高频和低频。对低频区域,采用开暗通道模型估计图像的透射率,并用引导滤波优化,得到清晰的低频和透射率图。对高频部分,结合透射率图并采用物理模型对信号增强。最后整合高低频信号重构出清晰无雾结果。对比单幅图像去雾,视频序列去雾清晰化的应用空间更为广阔但也更为困难。不同于单帧处理方案,本文提出一种保持时序一致性并有效减少视觉伪迹的实时视频去雾框架。基于视频相邻帧之间的相关性,定义时序一致性的透射率项。然后,融合当前透射率增量定义整体代价函数,权衡代价以求得最佳透射率图。同时为减轻低质图像恢复过程中存在的伪迹和噪声放大现象,提出利用相对无噪的有雾图像联合恢复潜在图像。实验结果表明,该提出框架能自适应消除了雾霾影响,并产生高质量的清晰视频。
[Abstract]:Foggy environment, random medium in the atmosphere, such as suspended particles, fog, haze and other effects of outdoor scene radiation visibility caused serious shortage, cause the camera to shoot pictures of existing equipment with low contrast, dim color degradation in different extent, affect people's visual sense of automatic extraction and useful information. Therefore, clear the restoration of degraded image has very important practical significance and application value. The method of defogging is insufficient in adaptability, and difficult to deal with different scenes and noise under the condition of fog degraded image and video. This paper analyzes the physical fog image degradation model, to explore in detail the key to image the fog technology used, the design and implementation of image and video defogging method. The main research results and contributions are as follows: according to the traditional based on dark channel to the region far from the viewpoint of color distortion, compared with Vignetter problem region near the proposed defogging method of scene depth segmentation and dark channel. Based on open plan is as follows: first detection input depth information of the image scene, divided the whole image into two parts and vistas; using the dark channel estimation model of atmospheric transmittance correction vision screen, regional value; fusion and vistas optimization of transmission ratio, and ultimately a clear fog image. The experimental results show that for different scenarios of the fog image, this method can effectively overcome the shortage of the dark channel prior method, visual effect has been significantly improved, adaptive. Observed frequency distribution in the fog reaching the image, combining with the method of defogging multi scale wavelet analysis. The image is transformed from the spatial domain to the frequency domain, the low frequency region to fog; in view of the fog on the high frequency information of the weakened, also need to high-frequency details with programs such as restoration. : first decompose the image into wavelet domain using wavelet image, high frequency and low frequency. In the low frequency region, using dark channel estimation model of image transmission, and guide the filter optimization, get clear frequency and transmittance map. The high frequency part, combined with the physical model of transmission ratio and finally integrated signal enhancement. The high frequency signal is reconstructed. Results no fog fog to clear contrast image space, application of video sequence to fog clear broader, but also more difficult. Different from the single frame processing scheme, this paper proposes a maintain sequence consistency and reduce the real-time video visual artifacts to fog correlation framework. Between adjacent video frames based on transmittance, the definition of temporal consistency. Then, the definition of the overall incremental transmission fusion cost function, in order to achieve the best balance of cost. At the same time to reduce the light transmission ratio is low In the process of quality image restoration, artifacts and noise amplification phenomena exist. It is proposed to restore latent images by using relatively non noisy foggy image. Experimental results show that the proposed framework can effectively eliminate haze effects and produce high-quality clear videos.

【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】