# underwater-image-enhancement

**Repository Path**: greitzmann/underwater-image-enhancement

## Basic Information

- **Project Name**: underwater-image-enhancement
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2019-10-09
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Image Enhancement
# Introduction
Visibility in underwater images is usually poor because of the attenuation of light in the water that causes low contrast and color variation. In this , a new approach for underwater image quality improvement is presented,based on [this paper](https://www.sciencedirect.com/science/article/pii/S2092678216302588/pdfft?md5=0fb801ccd8c3b98979cb631428a2c3e9&pid=1-s2.0-S2092678216302588-main.pdf). The proposed method aims to improve underwater image contrast, increase image details, and reduce noise by applying a new method of using contrast stretching to produce two different images with different contrasts.

# Mehodology
Here we are dividing a image from its average pixel value to highest & lowest respectively. Calling it as high contrast & low contrast respectively.Followed by this Rayleigh stretching of each pixel over its entire range. We get two images till now, we use their extremities to get an average image with averaging the noises.
The flow of process goes like this:

![Dummy Model Results](https://ars.els-cdn.com/content/image/1-s2.0-S2092678216302588-gr2.jpg)

# Input Image


![](https://github.com/ajit2704/im_enh/blob/master/ori.png)


1] Color channel decomposition
 ```python
    b,g,r=cv2.split(img2)
    b1,g1,r1=cv2.split(img1)
  ```

2] Finding average pixel
    
  First we split down the image in its respective b,g,r channel. We need to find average pixel for each channel.
  
  ```python
     imdb = ((b2.max()-b2.min())/2) + b2.min()
     imdg = ((g2.max()-g2.min())/2) + g2.min()
     imdr = ((r2.max()-r2.min())/2) + r2.min()

   ```
   
 3] Rayleigh stretching
     Its based on Rayleigh Distribution [algorithm](https://wikimedia.org/api/rest_v1/media/math/render/svg/8a06343fb89f74d188c25aef4931739d6c488467).
     Here using the basic principle of stretching the pixels over entire range,as mentioned in [eq2](https://www.sciencedirect.com/science/article/pii/S2092678216302588#formula2).
     Integrating this equation results in dynamic stretching of pixels over the entire range as shown in [eq4](https://www.sciencedirect.com/science/article/pii/S2092678216302588#formula4).
     Here Rayleigh stretching of 'b' channel for high contrast:
     
    
       b[b<imdb] = imdb
       for index,value in np.ndenumerate( b ):
           new_value = ((255* (value-imdb))/ ((b2.max()-b2.min())/(alpha**2)))
            b[index] = new_value
            
   4] Image composition
       We merges both set of channel to obtain high contrast & low contrast images:
     
     
        res=cv2.merge((b,g,r))
        res1=cv2.merge((b1,g1,r1))
    
   ![all text](https://github.com/ajit2704/im_enh/blob/master/con_screenshot_04.12.2017.png)
     
    
   5] Averaging two images, here we can use diffetent methods of averaging images, but the simplest for two images is weighted addition:
        
        res2= cv2.addWeighted(res,.5,res1,.5,0)

   6] Final stage is to increse the contrast of the image, there are no. of ways from which we can do this, like stretching s & v component image over entire range, gamma correction, histogram equalization.We need to maintain image for segmentation so we are omitting histogram stretching, instead we uses gamma correction:
        
     
        def adjust_gamma(image, gamma=1.0):
            invGamma = 1.0 / gamma
	          table = np.array([((i / 255.0) ** invGamma) * 255
		        for i in np.arange(0, 256)]).astype("uint8")
            return cv2.LUT(image, table)
       
     
  ![all text](https://github.com/ajit2704/im_enh/blob/master/avg.png)