The Research Of Image Dehazing Based On Dark Channel Prior

Image processing technology is more and more widely used in industry, medical treatment, education and other fields. Image processing system which is based on image feature extraction, such as outdoor video surveillance system, intelligent transportation system and automatic navigation system, is usually based on obtaining distinct image.However, under poor weather conditions, such as haze, rain, snow, images captured by camera sensor have lower contrast and color distortion, thus the images are blurred,bringing great problems to image processing systems in image processing and analysis.Therefore researching image restoration technology under poor weather conditions to enhance image contrast and resolution have great significance to guarrantee reliable and steady works of computer vision systems under poor weather condit Ions.This paper introduces variety image dehazing algorithms under haze conditions.Image dehazing algorithms are classified into two categories: nonmodel-based image enhancement algorithms and model-based image restoration algorithms. These nonmodel-based image dehazing algorithms don't take into account the attenuation process of the image and only enhance the contrast of the image to remove the influence of haze, thus these algorithms can improve the quality of image to some degree, but can not remove the influence of haze effectively. However, model-based image dehazing algorithms are based on the attenuation model of haze image to obtain the haze removed images, therefore can usually achieve better dehazing effects. For nonmodel-based image dehazing algorithms this paper introduces these algorithms: global histogram equalization algorithm, local histogram equalization algorithm, Retinex algorithm and homomorphic filter algorithm. Local histogram equalization algorithm includes patch overlapped local histogram equalization algorithm, patch non-overlapped local histogram equalization algorithm and patch partially overlapped local histogram equalization algorithm. Patch partially overlapped local histogram equalization algorithm reaches a compromise between dehazing quality and dehazing speed, thus having superiority in image dehazing over the other two algorithms. Retinex algorithm is classified into Single-Scale Retinex algorithm, Multi-Scale Retinex algorithm,Multi-Scale Retinex with Color Restoration algorithm and iteration based Retinex algorithm. Model-based image dehazing algorithms include depth-based algorithm,partial differential equation based algorithm and prior-based algorithm. This paper mainly introduces prior-based algorithm, such as Tan's algorithm, Fattal's algorithm,Tarel's algorithm and He's algorithm.Based on He's dark channel prior dehazing algorithm, this paper proposes a novel improved dehazing algorithm.Firstly, this paper transforms the color space of dark channel prior from RGB to YCbCr. Then this paper ultilizes min filter only on the Y component to estimate the initial transmission. This modification can greatly accelerate the dehazing speed while guaranteeing the dehazing quality. Secondly, this paper introduces two kinds of filter which are aimed to optimize the transmission: soft matting and guided filter. Based on the guided filter, this paper proposes a fast guided filter using bicubic interpolation. This modification again accelerates the dehazing speed while guaranteeing the dehazing quality. Finally, aiming at the invalid of dark channel prior in sky region, which results in color distortion and halo phenomenon, this paper proposes an algorithm to improve the dehazing effect in sky region. This paper compares our algorithm with He's algorithm and other prior-based algorithms in dehazing quality and dehazing speed, then arrive at the conclusion.