基于边缘保持滤波的单幅图像快速去雾
A Fast Haze Removing Algorithm of Single Image Using Edge Preserving Filtering

为了解决基于暗通道先验的图像去雾算法运行效率低下的问题以及天空等明亮灰白区域去雾后的色彩失真问题,提出一种基于边缘保持滤波的单幅图像快速去雾算法。首先根据暗通道先验规律,得到粗略的透射率图和大气光估计值;然后用边缘保持滤波算法对粗略透射率滤波得到细节平滑、轮廓清晰的精细透射率图;再用阈值法对灰白明亮区域的透射率修正,之后用边缘保持滤波算法对修正后的透射率进行平滑,得到最终的透射率图。根据估计的大气光和透射率,利用大气散射模型即可恢复出无雾图像。经测试,该算法不仅具有很高的运行效率,而且对各种类型的薄雾图像都有较好的去雾效果。客观评测也表明,该算法在对比度增强程度、色调还原程度、结构信息复原程度方面的综合指标都优于其他算法。另外,所提算法还能够实现图像处理器(GPU)像素级的并行运算,对于分辨率为1 280像素×1 024像素的彩色图像,用型号为NVIDIA GeForce 9 800 GT的GPU处理,速度可达10帧/s。
A fast haze removing algorithm of single image is proposed to solve the problem that the image dehazing algorithm using dark channel prior is of low efficiency and the color distortion of the bright grey area happens after dehazing. The algorithm is based on the edge preserving filtering, and first gives rough estimations of the transmittance and the atmospheric light using dark channel prior. Then, these rough transmittances are optimized through using the edge preserving filtering algorithm to get refined transmittances with smooth details and clear outlines. The bright and grey areas in the refined transmittances are corrected using a threshold, and the final transmittance image is generated by using the edge preserving filtering once more to smooth the corrected transmittances. Lastly, the atmospheric scattering model is used to recover the haze image from the estimated atmospheric light and the final transmittance image. Test results show that the proposed algorithm not only has the very high efficiency, but also has a preferable effect in dehazing all kinds of images with thin haze. It follows from the objective evaluation that the algorithm is superior to other existing algorithms in the aggregative indicators including contrast enhancement, color reduction and structural information recovery. Moreover, the proposed algorithm can be realized in pix？？level parallel computation using GPU. When the NVIDIA GeForce 9 800 GT GPU is used, the processing speed reaches 10 frames per second for a range of 1 280*1 024 resolutions