Avoiding Overfitting: A Survey on Regularization Methods for Convolutional Neural Networks

Several image processing tasks, such as classification and object detection, have been significantly improved using Convolutional Neural Networks (CNN). Like ResNet EfficientNet, many architectures achieved outstanding results in at least one dataset by the time of their creation. A critical factor training concerns network's regularization, which prevents structure from overfitting. This work analyzes several regularization methods developed last few years, showing significant improvements for different CNN models. The works are classified into three main areas: first is called "data augmentation", where all techniques focus on performing changes input data. second, named "internal changes", aims to describe procedures modify feature maps generated neural network or kernels. one, "label", transforming labels a given input. presents two differences comparing other available surveys about regularization: (i) papers gathered manuscript, not older than five (ii) second distinction reproducibility, i.e., refered here code public repositories they directly implemented some framework, TensorFlow Torch.