Evolution of image filters on graphics processor units using Cartesian Genetic Programming

Abstract

Graphics processor units are fast, inexpensive parallel computing devices. Recently there has been great interest in harnessing this power for various types of scientific computation, including genetic programming. In previous work, we have shown that using the graphics processor provides dramatic speed improvements over a standard CPU in the context of fitness evaluation. In this work, we use Cartesian Genetic Programming to generate shader programs that implement image filter operations. Using the GPU, we can rapidly apply these programs to each pixel in an image and evaluate the performance of a given filter. We show that we can successfully evolve noise removal filters that produce better image quality than a standard median filter.

Cartesian Genetic Programming for Image Processing

“…Several key examples also use CGP (Harding, 2008;Smith et al, 2005;Vasicek and Sekanina, 2007;Mart « "nek and Sekanina, 2005), and these examples all use mathematical and logical operators to define a simple convolution operation. Here we show CGP-IP addressing both 'salt and pepper' and Gaussian noise.…”

“…shift takes two parameters to say how far to shift in the horizontal and vertical directions. Shifting is an important feature, as it allows for processing of the neighbourhood (see (Harding, 2008) for details). reScale, downsamples the image by Parameter0 factor and then upscales it again to the original size.…”

“…At each pixel location, an evolved expression takes the in neighboring pixels' values, and computes a new value for the centre pixel (Gonzalez and Woods, 2006). Typically these programs operate at a mathematical level, where operations such as +, −, × and ÷ are used (Harding, 2008). Implementations on Field Programmable Gate Arrays (FPGAs) for noise reduction use a mixture of binary operations (e.g.…”

Parallel Genetic Programming on Graphics Processing Units

“…In contrast to [16][17][18] where the candidate programs were compiled to execute on GPUs, [28] showed a way of interpreting the trees. While the previous presented approach requires that programs are large and run many times to compensate the cost of compilation and transference to the GPU, the interpretable proposal of [28] seems to be more consistent because it achieved speed ups of more than an order of magnitude in the Mackey-Glass time series and protein prediction problems, even for small programs and few test cases.…”

“…Following the same idea of compiling the candidate solutions, [16] uses a Cartesian GP on GPU to remove noise in images. Different types of noise were artificially introduced into a set of figures and performance analyses concluded that this sort of parallelism is indicated for larger images.…”