CNN-Based Retinex Technology

Huang Chao-Hui, and Lin Chin-Teng


--------------------------------------------------------------------------------

Introduction

We introduce a method to implement Retinex Technology on Cellular Neural Networks (CNN). Retinex is an adaptable algorithm which can enhance the image as human-vision did. In human-vision, As well known the most important thing is the silhouette of objects. Retinex can enhance the edge or contour of objects in vision sense. Even aggrandize that. Thus, the objects in image will become to more apparent.

On other hand, Retinex also can compensate the prenatal deficiency of human-vision. Thus, Retinex make some important but unobvious information may not be lost as well as possible.

CNN is another interesting concept, which is an algorithm as powerful as human-brain. Basically, CNN is similar to the action of neurons. Via the connection between several cells, we may able to implement some difficult tasks.

Both Retinex and CNN will be introduced as followed sections.

This algorithm have been published on European Conference on Circuit Theory and Design 2003 (ECCTD '03). If you need detail information, please click here.


--------------------------------------------------------------------------------

Experiment

We tested this algorithm both on static image and video. One of well-known benchmark of Retinex Technology is a X-Ray image demonstration. We have tested our algorithm on this image. The result is as following.

X-Ray Image Enhancement


Original Image

Iteration Time = 1 (Tao)


Iteration Time = 2 (Tao)

Iteration Time = 4 (Tao)


General Image (Lenna) Enhancement


Original Image

Iteration Time = 1 (Tao)


Iteration Time = 2 (Tao)

Iteration Time = 4 (Tao)




Also, we tested our algorithm on video. The result is as following.

Video Enhancement


Original Video

Video after Processing



--------------------------------------------------------------------------------

Conclusion

Under the architecture of this algorithm, we have obtained many good results. However, in video case, there are some ˇ§flickingˇ¨ phenomenon appeared. We are still tried to solve this problem.


--------------------------------------------------------------------------------