Article Preview
TopIntroduction
Global warming and pollution are on the rise due to exponential technological growth. As a result, atmosphere is becoming contaminated with APM (Atmospheric Particulate Matter) in the form of fog, dust, vog, industrial waste, etc. Impact of APM leads to decrease in visibility (Mao, 2015). Atmospheric effect and inherent camera distortion in outdoor image make detection of distant objects difficult and captured images become inadequate in luminance, contrast, brightness and colour. Therefore, visibility improvement, and colour restorations in computer vision applications are very important. In this research work atmospheric scattering based optical image formation and the retrieved model developed by H. Kosmeider in 1924 and revisited by E.J. McCartney 1976 has been studied and analysed (Koschmieder, 1924, & McCartney, 1976). Single image dehazing generates good quality image using DCP(Dark Channel Prior). Fixed value of haziness factor ω (0.95) and the sky area or bright areas are two drawbacks of the DCP algorithm. This method is efficient, but slow with high time complexity and the output image appears dull (He et al., 2009). The work of R Tan deals with gray or colour image. Data cost and smoothness cost are extrapolated from airlight in the framework of Markov Random Fields optimized by Graph-cut or belief propagation. The method is efficient, but not applicable for real time (Tan, 2008). The algorithm proposed by J.P. Tarel is fast and its complexity is a linear function with the number of image pixels for both colour and gray image. The algorithm is tuned by only four parameters, atmospheric veil inference, image restoration, smoothing, and tone mapping (Tarel et al., 2009). Research work of R. Fattal is based on new transmission estimation that develops refined image formation model where surface shading and transmission function are found to be locally statistically uncorrelated. The result of this effect removes the scattering light and resolves the ambiguity in hidden pixel information to retrieve clear image with original scene colour and contrast (Fattal, 2008). Berman and Avidan emphasised a method based on non-local prior. Uneven degradation depends on the distance between camera and the pixel. Transmission coefficients attenuate every pixel and modulate the amount of haze with the help of haze line to estimate the amount of haze for original haze-free image. The algorithm is linear, fast, deterministic with no training required, but still cannot applicable for real-time applications (Berman & Avidan, 2016). The work based on colour attenuation prior is a linear model with supervised learning. It is robust and efficient. Although it is fast, this method is also not suitable for real time use. The output images are little bit whitish, and scene depth is less (Zhu et al., 2015). Authors are working on fast weather independent visibility improvement algorithm which can be applied in surveillance, satellite image processing, car driving, aircraft landing system, etc. (Roy & Chaudhuri, 2016). Original image geometric structure has been recovered optimally from the joint estimation of scene albedo and depth of single foggy image using Bayesian statistical method. Here, image is modelled with Factor Markov Random Field where scene albedo and depth are considered statically independent and estimated jointly. Natural images and depth statistics are being used for scene albedo and depth estimation for foggy images with canonical maximisation algorithm having alternate minimisation. The method is effective and efficient, but having high computational complexity making the algorithm not suitable for fast applications (Nishino et al., 2012). Inherent boundary constrains on transmission estimation has been solved with regularisation method. Weighted L1-Norm contextual regularisation has been proposed for optimal transmission estimation of single deep unclear image. Apart from that, variable splitting algorithm has also been studied. Good quality, high colour fidelity, fine detailed image can be obtained with few general assumption (Meng et al., 2013). DehazeNet (Trainable end-to-end system) has been used for transmission estimation. DehazeNet is CNN (Convolution Neural Network) with deep architecture whose hidden layers are designed to remove haze from single image. Maxout, Pooling layers, are designed to extract haze relevant features for generating haze free output. Bilateral Rectified Linear Unit (BReLU) is the key to generate optimum haze free image from single image. The method is highly effective only for off-line process. (Cai et al., 2016)