FPGA based Computing Acceleration of DDE
Mentor : Liu Ning
Member : Wu Jinzhi, Men Xiaolang, Qi Changjun
DDE is the abbreviation of Digital Detail Enhancement, a concept that improves image details through digital algorithms and a technology that is widely used in the field of image processing.This work based on FPGA platform aims at solving problem about narrow histogram distribution of thermal imaging and lack of detail in traditional processing methods.
The first picture is the original picture sampled from thermal imager. Against most people's impressions, it's so blurred that it's almost impossible to distinguish for its narrow histogram distribution. The second picture, thermal images with commercial general processing, is processing by histogram equalization, which has a fatal flaw in image’s detail───Histogram equalization can't enhance image details, and even lose details. The last one is the effect of method DDE(Digital Detail Enhancement).
In order to improve the above two problems, we recommend the use of method DDE.But this method also has its drawbacks. First, the complex operations are difficult to achieve real-time performance, even on high-performance CPUs. Second, the cost of using GPU is too high to be commercially available. My team used one year to complete the design of the processing unit based on FPGA, and the final effect was excellent.
The pipeline architecture designed based on FPGA is as above. We constructing algorithms based on thought that separating an image into high and low frequencies by a low-pass filter(Gauss filter), and then processed separately.
The work processes low-frequency video with Histogram equalization to overcome the problem of narrow histogram distribution in thermal imaging and processes high-frequency video with adaptive detail enhancement to overcome the lack of detail in existing processing methods.
Finally, the processed high-frequency and low-frequency video streams are fused again, which realizes DDE.The work achieves excellent real-time performance and will not cause any frame loss during the real-time processing.
The picture above is a demonstration of the work's indoor scene affairs. As shown in the displays in the first two pictures, the upper left corner is the original video, the upper right corner is the DDE effect, the lower left corner is the DDE detail layer, and the lower right corner is the traditional processing effect.
From the details of hairs and clothes wrinkles in the first figure and the inscriptions on the sword in the second picture, we can see that DDE algorithm is far superior to the traditional algorithm.