CuNeuQuant: A CUDA Implementation of the NeuQuant Image Quantization Algorithm
David Bottisti, Liuva Mendez, Damian Dechev · Journal of International Crisis and Risk Communication Research · 2012
Color quantization is an often performed pre-step in many image processing and computer vision applications. Quantization is defined as the process of selecting a palette of representative colors P which can replace the original colors C in an image such that |P| ≪ |C| and the perceptual distortion of the reduced color image is minimized. It is well known that the quantization process is an NP-complete problem and as such, many competing heuristic algorithms exist. One high-quality quantization algorithm is NeuQuant due to Dekker. In this paper, we describe a GPU based parallel implementation of the NeuQuant algorithm. Our GPU-based approach demonstrated a speedup by a factor of 5 or more in the performance evaluation we have performed. The details of the NeuQuant algorithm present unique difficulties to implementing a parallel version due to the sequential dependencies present when training the underlying neural network.