Fog Image Enhancement Algorithm Research Based On Local Information Feature

Fog, the air filled with a large number of small droplets, natural light will be generated by scattering, enabling the human eye or optical sensor to receive a blurred visual image information, so fog reduced the visibility of the atmosphere, affected outdoor activities seriously such as the transportation, video surveillance, target tracking and even the security and reliability of military and other important areas. This thesis studies the fog image enhancement algorithms, and proposes new algorithms by analyzing the shortcomings of existing algorithms to improve the clarity of the image fog.At present, the fog image enhancement algorithms can be classified into two types: spatial-domain and frequency-domain algorithms. This thesis proposes some improved algorithms according to the characteristics of fog image. The main contents are as follows:1. After the traditional histogram equalization processes fog image, there is enhanced unevenly shortcoming. This thesis analyzes the reasons for this defect and proposes a new localized histogram equalization algorithm by combining the brightness preservation histogram equalization algorithm and the interpolation algorithm. This new algorithm will get distinct local details and also to maintain the brightness. The experimental analysis shows that: The new algorithm is superior to the traditional histogram-based enhancement algorithms in the details enhancement, brightness preservation and the overall effect.2. Because of the fixed filter the existing center/surround Retinex algorithm can't enhance different level fog or multi-depth images effectively in details enhancement while maintaining the color fidelity. So a new Retinex algorithm based on variable filter is proposed to improve the existing Retinex algorithm. First the thresholds are worked out according to the image quality distributing ; Then in each sub-block, the algorithm uses the local information and the thresholds to make a corresponding filter to estimate illumination of the current sub-block, and adopt the approach of partially overlapped sub-block to get the whole image's illumination ; At last the whole illumination is subtracted from original image. The experimental results show that: The proposed algorithm can enhance the multi-depth or different level fog image effectively. 3. The traditional Multi-Scale Retinex method can't enhance the details effectively. This thesis proposes a improve Multi-Scale Retinex by adopting information fusion strategy based on wavelet transform domain to replace the linear weighted average. At first, the basic strategy of image fusion is taking four times wavelet decomposition; Then, the high-frequency take absolute maximum to enhance the details and the low-frequency adopt the strategy which based on local variance to modify color. The subjective and objective evaluation show: This new MSR algorithm is better than the traditional algorithm in details and color fidelity.