Beijing, December 14, 2012
A real-time image defogging technology has been developed in Fujitsu Research & Development Center Co., Ltd. (Note1). It can process images degraded by fog, haze, dust, and air pollution, and get restored images clearly in a high speed.
Compared with existing method, its run speed improves about 100 times. Images with resolution of 720x480 can be processed in real-time with software on PC. At the same time, this technology features low computational complexity and low memory demand. It can be fused into surveillance cameras with low complexity hardware implementation.
【Background】
All types of cameras, such as digital cameras, surveillance cameras, and car-mounted cameras, have been applied in many aspects of life, with the advances in technology and improvement of living. However, bad weather (fog, haze, dust, rain, and snow, etc.) will reduce the performance of outdoor cameras, and make them hard to keep stable visibility. For surveillance cameras and car cameras, the images will be blurred under the situations of heavy fog, dust, and air pollution. People cannot identify clearly moving cars, pedestrian, and car plate number. To solve such problems, we have developed a fast high performance defogging technology.
【Topics】
Two types of technologies exist for improving visibility in foggy days:
1. (fog correction) adjust brightness and contrast for foggy images
•High speed, low image quality
2. (fog removal) estimate fog density for each pixel, then remove the fog
•Good image quality, low speed.
Our technology is based on fog removal. The basic idea of fog removal is, foggy image = clear image + fog density. That means foggy images are clear images polluted by fog. If we estimate the fog density and remove it, then the clear image can be got.
In order to get high quality defogged image, the way of fog removal needs to get optimized value for each pixel. Therefore a complex filter with a large size is needed. If implemented with software, it is slow because of high complexity. If implemented with hardware, the computational complexity and memory requirement is very high. Our research topic is to reduce the computational complexity and memory demand while keeping high performance.
Figure 1. Image defogging model
【Technology】
The procedure of defogging is shown as follows.
In traditional methods, the step of getting transmission map will cost much time. We have developed a fast method for this.
• fast defogging method
A kind of fast high-quality defogging method, which calculates two transform maps (coarse/fine) from the input image, and then combines them to do local filtering for each pixel.
The atmospheric light is estimated as follows:
• we think the color of fog and haze is consistent, without color offset. However, the colors of natural objects are different.
• we think that the farthest part of background is the area where the fog density is highest, and the place has the best consistency is the sky.
Figure 2. Estimate atmospheric light
Figure 3. Get transmission map
【Result】
(1) High quality, real time processing with software implementation
(2) Processing result
【Future】
We will test and improve our algorithm in field, and get real applications in 2013. At the same time, we will consider to implement it with hardware, and make possible fusion with surveillance cameras and in-vehicle cameras, to achieve broader applications.
【Note】
(Note1) Fujitsu Research & Development Center Co., Ltd.: Chairman Shigeru Sasaki. Location: Beijing, China