Study On The Object-oriented Image Defogging Using Dark Channel Prior

In foggy days, due to the absorption and scattering of light from suspended aerosol particles in atmosphere, images received by the camera usually have poor contrast and color fidelity, details of the information is lost, many features are covered or attenuated, which bring great difficulties to the subsequent image processing and analysis. Removing fog from a degraded image and enhancing its sharpness and contrast, which improves the reliability and stability of computer vision system in bad weather, will have very important theoretical significance and practical value.The image defogging using dark channel prior is one of the defogging methods based on physical imaging model, which has been widely used due to its reasonable physical interpretation and better restored results. For some drawbacks of the approach, this paper made some corresponding improvement studies. Which can be described as follows:(1) with the object-oriented method, divide the image into many homogeneous regions by using the algorithm of watershed segmentation, and use the minimum filter in every block to get its dark channel, which leads to eliminating the “halo effect” existing at the edge area with abrupt depth jumps effectively;(2) Obtain a more accurate atmospheric light according to extracting the sky area of the image;(3) use the method of “guided filter” to smooth the transmission map while remaining its sharp edges, and enhance the running speed of this algorithm;(4) propose a new adaptive bilinear tensile method for restored image brightness adjustment, overcome the shortcoming of its dark vision.For the outdoor computer vision system, the foggy video needs to be processed in real-time, but the single-core computing capability of CPU cannot support this requirement. In order to improve the running speed of the algorithm approached in this paper, we take GPU as the parallel processing hardware platform, use CUDA as the programming model, implement and optimize the parallel object-oriented defogging algorithm using dark channel prior, finally, we can process the standard definition video with resolution 720*576 at the speed of 25 frames per second by our parallel algorithm.