Optimizing Convolutional Neural Networks for Stereo Matching by using Multi-Objective Evolutionary Algorithms
Ali Zoljodi
Computer vision is one of the most applied sciences related to the AI. One of the computer vision applications is extracting three-dimensional scenes from stereo pictures by using convolutional deep neural networks. The problem of these networks especially in this application is their cost of processing which decreases by losing the accuracy rate in the network. The methods for optimizing these networks are generally manual which need much knowledge and effect by human mistakes. In this research, we try to utilize automatic methods like evolutionary algorithms to decrease the network's cost of processing besides preserving the current accuracy percent. To achieve this goal we optimized a piece of one of the most famous stereo matching networks by using a simulated annealing algorithm. The output of this method is saving the accuracy rate besides reducing processing costs about 22 percent.