The Impact of Input Image Data Size on The Training Speed of Convolutional Neural Networks

In using neural networks for classification tasks, the most time-consuming process is repeated training of the network. Due to the development and progress of hardware equipment may only take a few hours for a model that required hundreds of hours to train in the past [1]. To obtain more accurate results or increase prediction speed, multiple repeated experiments are often carried out during the experiment. Due to the different performance of the equipment owned by actual users, how to increase the training speed of the neural network as much as possible without affecting the final accuracy? In the process of using neural networks for classification tasks, especially in learning how to use neural networks, there is a certain importance in the network process. The main purpose of this work is to study the impact of simplified input data on training duration when training on low-performance execution devices while affecting execution performance as little as possible. The effect of reducing the amount of data in a certain dimension on the training time and results were compared during the experiment. This work shows that replacing colour images with grey images for text recognition training can effectively reduce the length of network training, and the impact on accuracy can be ignored in most cases. The results show that reducing the amount of data in the picture when the accuracy requirements are not very strict, although it will cause a certain amount of information loss, has a significant effect on shortening the training time.