A Perspective of the Noise Removal for Faster Neural Network Training
Martin Rajnoha, Vojtech Mikulec, Radim Bürget, Jiri Drazil · 2019
Image classification is widely used within image processing area. It is known that insufficient amount of data has negative impact on the training of neural networks in terms of accuracy, convergence speed and in some cases even in the inability to converge. On the other hand, big amount of data significantly increases the training time and costs needed for model creation. Every training sample contains the part valuable for decision (face in case of this paper) and noise, i.e. background of the object. This paper introduces method of iterative noise removal during the training with combination with the transfer learning to optimize the speed of the training process. We show the combination of proposed noise removal and transfer learning leads to more effective training process and enables to learn also from limited data sets. The main contribution of this paper is a proposed method that reduces training time and it is able to accelerate the process in average by 69%. The method was tested on binary classification of two persons from LFW database.