An Efficient Description with Halide for IIR Gaussian Filter

Recursive Gaussian filter is one of the constant-time algorithms fir Gaussian filtering to reduce the computational order. The recursive property makes programs hard for optimization because all resulting pixels in a recursive filter depend on all input pixels. Halide is a domain-specific language for image processing and can powerfully accelerate image processing with a concise description. However, it is also tricky to describe concise scheduling for recursive filters. As a solution to this problem, RecFilter has been developed as a domain-specific language for handling recursive filter, which internally generates the Halide code. The limitation of RecFilter is in approximation accuracy. Here, we extend RecFilter to suit the methods commonly used in recursive Gaussian filter to solve this problem. Also, we improve computational performance by changing computational scheduling. Experimental results showed that the proposed generator produces more computationally efficient codes, and the resulting images have higher approximation accuracy.