Image Deraining

The rain components are removed from the image based on the rain characteristics. The coloured image is divided into high frequency and low frequency parts so that the high frequency part consists of most of the rain components. Then by using Dictionary learning method the rain components are extracted from the high frequeny part.
To extract more non-rain details we use Sensitivity of variance of color channels(SVCC). Finally the non rain component part and low frequency part are combined to get the image without rain.

The visual quality of images is mostly affected by the weather conditions like snow, rain etc. In such cases Rain removal is an important aspect since the rain images have a serious impact on many computer vision algorithms like object recognition and detection recognition, detection, tracking etc .

Here we have employed rain removal from a single image since for industrial and academic purpose it is more flexible. In this paper, we performed rain removal from a single color image based on analysis of characteristics of rain pixels. Therefore, we make a brief summary on the simple but quite useful characteristics of the rain ?rst.
Firstly, all the rain pixels fall in the high frequency part of an image since rain reflects light stronger than any other particles.
Secondly,the rain streaks and other particles are distinguished based on the fact that is there often exists an edge jump between them. Therefore, an image containing rain streaks will have high average Horizontal Gradient.

The rain pixels appear in the constant areas of an image and their value doesn’t change much after applying filter .Therefore the background intensity is taken as the value of the rain pixel in low frequency part and the corresponding value in the high frequency is the change in intensity after being affected by rain. Iorig =Ilf + Ihf.
The algorithm which is shown beside is used to extract the non rain components from a rain image in pixel domain.

At last the components which are free from the rain components are combined together to get the final image which is free from rain i.e; Ifinal = Ilf +HF nr1+HF nr2 +HF nr3
Based on the fact that the rain pixels reflect light stronger than other pixels we can roughly estimate the position of the rain pixels in the image. For a normalized image say I we need to calculate mean values for each pixel i.eI (x,y) is pixel position and 5 mean values be Ij (j=1,2,3,4,5).
A window Wj of suitable size must be selected and the mean values must be calculated with the pixel I(x,y) being at centre,bottom-left, top-left, bottom-right, top-right. Then if the below equation becomes true for every j value then the corresponding pixel i.e I(x,y) is regarded as rain pixel.

A location matrix of size equal to I is taken and all the rain pixels are made 0 at corresponding rain pixel postion (x,y) and taking 1’s at remaining pixel postions.
By this the location matrix only contains only 0’s and 1’s. Now the original image I is multiplied with the location matrix L(multiplication is done pixel to pixel i.eScalar multiplication) so that all the rain pixels are made zero and the resulting image is given to bilateral filter which separates out the low frequency part of the image .Then the High frequency part is calculated simply by subtracting the low frequency part from the original image i.e Ihf = Ioriginal – Ilf
To extract the non rain components present in the high frequency part of image, we use the dictionary learning method which uses the morphology component analysis to represent the high frequency part in a thinly dispersed manner.

We know from the second characteristic of rain that the colour channel variance of non rain pixels is higher when compared to the rain pixels. So, from the dictionary learning process we will find the sum of variances of all atoms present in the high frequency part.
For rain atoms the colour variance of rain atoms is very small and it is nearly equal to zero. Therefore by fixing a threshold parameter (/1) we separate the non rain component HFnr1 and the rain component HFr1 from the obtained high frequency part.

After the above process, some of the non rain pixels whose colour intensity values are nearly same to that of rain pixels still exist in high frequency part. So by using another characteristic of rain which is based on horizontal gradient is used for further separation. We will calculate the horizontal gradient of every pixel in a dictionary atom, and fix a threshold parameter (/2).All the pixels which are having less horizontal gradient than threshold are considered as non rain pixels (HFnr2) and those which are having higher value than the threshold are considered as rain components.
The above obtained non rain components i.e.,HFnr1 and HFnr2 and the low frequency part of the original image are added together to form the final rain free image. Ifinal = Ilf +HF nr1+HF nr2 +HF nr3