Evolving binary-weights neural network using hybrid optimization algorithm for color space conversion

Abstract

Arti cial Neural Networks (ANNs) are applied to many complex real-world problems, ranging from image recognition to autonomous robot control. However, to design a neural network that can implement special task, it is necessary to select an appropriate biological neuron model, meanwhile, good learning algorithm should be adopted to achiee the expected goal. euroevolution is a form of machine learning that uses Evolutionary Algorithms (EAs) to train ANNs. EAs, for the learning algorithm used by neural networks, can provide alternative and complementary solution, which can avoid the frequently happened issues of getting stuck in local minimum" during the iteration process made by gradient-based learning algorithms. In this paper, a method using Hybrid PSO-based Learning Algorithm (HPLA) to evolve the connection weights and network parameters of Binary-Weights Neural Network (BWNN) will be introduced. The extracted knowledge from trained BWNN can then be used to construct high-speed shift-and-add based Color Space Converter (CSC) hardware architecture. The experimental results in this research also show that the performance of implemented hardware architecture is good at RGB to YCbCr color space converting, and it also has the advantages of high-speed and lowcomplexity.