TECHNION - ISRAEL INSTITUTE of TECHNOLOGY Department of Electrical Engineering The Vision Research and Image Science Laboratory

Abstract

Usually there are differences between a recorded color image and the direct observation of the same scene. One of the big differences is the fact that the human visual system is able to distinguish details and vivid colors in shadows and in scenes that contain illuminant shifts.

In this project we examine the performance of an algorithm called A Multiscale Retinex with Color Restoration (MSRCR), presented in [1]. This algorithm tries to imitate human visual “computation” while observing scenes that contains lighting variations. MSRCR is an extension of a former algorithm called Single Scale center/surround Retinex (SSR) [2]. MSRCR achieves simultaneous dynamic range compression, color consistency and lightness rendition. In this project we look for the optimal value of each parameter of the SSR and the MSRCR. The MSRCR algorithm was compared to other two, well known, image enhancement methods – gamma correction and histogram equalization. Results indicated that the performance of the Multiscale Retinex is consistently good, while the performance of the others is quite variable.

Basic Approach

The general mathematical formulation of the center/surround Retinex is

where  denotes the Retinex output, I_i(x,y) the image distribution in the i'th color spectral band, “*” the convolution operation, and F(x,y) the surround function

where c is the Gaussian surround space constant and K is selected such that

The MSR output is then simply a weighted sum of the outputs of several different SSR output.

where N is the number of scales,  the i'th component of the n’th scale, the i'th spectral component of the MSR output, and w_n the weight associated with the n’th scale. The only difference between and is that the surround function is now given by

The color restoration method for the MSR is given by

where is a constant parameter of the color restoration function. The MSRCR is given by

The final version of MSRCR can be written as

where G and b are the final gain and offset values respectively.

1. 1. Daniel J. Jobson, Zia-ur Rahman, Glenn A. Woodell; “A Multiscale Retinex for Bridging the Gap Between Color Images and the Human Observation of Scenes”, IEEE Transaction on Image Processing, vol. 6, no. 7, July 1997.

In this project we examine several aspects concerning the SSR:

1. the form of the surround function
2. the placement of the log function
3. the scale constant for the surround function
4. the treatment of the Retinex result before display

1. the combination three scales (small, intermediate and large) of the SSR functions
2. the weights associated with the scales
3. the color restoration constant.

The algorithm was realized as a MATLAB program. In order to ease the research and to make it more convenient to get the visual results, we developed a graphic user interface (GUI), see Figure 1. The GUI allows the user to change the MSRCR constants before running the algorithm and to compare the resulting performance to the performance of a standard brightness correction (gamma correction).
 

Results

Figure 2 illustrates the MSRCR performance. As seen the MSRCR algorithm achieves graceful dynamic range compression together with tonal and color rendition. The MSRCR algorithm was also tested on some images with familiar colors and no strong lighting defects and the results showed that the MSRCR does not introduce significant distortions into images without lighting variations.
 
 
 

Conclusions

1. The performance of the three surround functions: inverse square, exponential and Gaussian are quite similar. This result stands in contrast to the results presented in [2], in which the Gaussian’s performance was significantly better. The Gaussian function was chosen as the form of the surround function because of compliance considerations and due to its small width in comparison to the other two.
2. The placement of the log function should be after surround formation.
3. The scale constant of the Gaussian function should be  in order to get a satisfying image quality. This range of C₃ is quite close to the range presented in [2] ().
4. The MSR is composed of three SSR functions with small(c=80), intermediate (c=120) and large (c=250) scale-constants. This combination allows the MSR to synthesize dynamic range compression, color consistency and tonal rendition, except for scenes that contain violation of the gray-world assumption. The color restoration function produces good color rendition even when a gray-world violation occurs. The scale constants that were chosen in [1] are: c=15, c=80, c=250. We chose larger constant values because small values produce desaturation of colors in the final image.
5. The Retinex output has to be processed before displaying. This includes clipping some of the lowest values in order to get better color rendition. The gain and offset constants do not vary either from image to image or between color bands.

Acknowledgments

We would like to thank our supervisor Dr. Doron Shaked for his support and guidance throughout this project. Also we would like to thank Johanan Erez for his attention and help during the whole project and the Ollendorff Research Center Fund which supported this project.

Additional documents