Block matching noise reduction method for photographic images applied in Bayer RAW domain and optimized for real-time implementation

Abstract

Image de-noising has been a well studied problem in the field of digital image processing. However there are a number of problems, preventing state-of-the-art algorithms finding their way to practical implementations. In our research we have solved these issues with an implementation of a practical de-noising algorithm. In order of importance: firstly we have designed a robust algorithm, tackling different kinds of nose in a very wide range of signal to noise ratios, secondly in our algorithm we tried to achieve natural looking processed images and to avoid unnatural looking artifacts, thirdly we have designed the algorithm to be suitable for implementation in commercial grade FPGA's capable of processing full HD (1920×1080) video data in real time (60 frame per second). The main challenge for the use of noise reduction algorithms in photo and video applications is the compromise between the efficiency of the algorithm (amount of PSNR improvement), loss of details, appearance of artifacts and the complexity of the algorithm (and consequentially the cost of integration). In photo and video applications it is very important that the residual noise and artifacts produced by the noise reduction algorithm should look natural and do not distract aesthetically. Our proposed algorithm does not produce artificially looking defects found in existing state-of-theart algorithms. In our research, we propose a robust and fast non-local de-noising algorithm. The algorithm is based on a Laplacian pyramid. The advantage of this approach is the ability to build noise reduction algorithms with a very large effective kernel. In our experiments effective kernel sizes as big as 127×127 pixels were used in some cases, which only required 4 scales. This size of a kernel was required to perform noise reduction for the images taken with a DSLR camera. Taking into account the achievable improvement in PSNR (on the level of the best known noise reduction techniques) and low algorithmic complexity, enabling its practical use in commercial photo, video applications, the results of our research can be very valuable.