Review of Convolutional Neural Network Optimization and Training in Image Processing

Major breakthroughs in the fields of image processing, pattern recognition, and scene classification have recently been made through the using of convolutional neural networks (CNNs) and deep learning. If the training set is sufficient, a CNN performs better than traditional machine learning algorithms and differs from them in many ways. In CNNs, feature extraction is more intelligent, so researchers no longer require extensive knowledge about the specific topic. They can focus on the CNN itself, including the structural design of the network, the model optimization, and the numerical solution. The many articles on CNNs published over the past few years propose many neural network models. Methods of training and optimization for CNNs have also been proposed. This paper reviews the history of CNNs, introduces the commonly established CNN structure, and summarizes methods and tips for CNN training.