Artwork recognition based on ResNet models with an added batch normalization layer

Artwork recognition is an important and unique application of image recognition networks. Advanced art recognition networks can not only help to construct digital archives for artwork, but also serve as educational tools for people to use to learn about art. Compared to other types of image recognition tasks, artwork recognition has unique challenges. Not only can different artists have similar artistic styles, but artwork from the same artist can also differ significantly, making it difficult to categorize artwork. Convolutional Neural Networks have always been used to perform image recognition tasks. This research proposes two ResNet50 and ResNet152 models with an added batch normalization layer. Many algorithms that mitigate overfitting during training processes are also explored in this paper. The ResNet152 model outperformed the ResNet50 model, reaching a validation set accuracy of 88% and a test set accuracy of 86%. Compared to different artwork recognition models in previous researches, the network introduced in this paper not only has a simpler architecture, but also generalizes better on a larger comprehensive dataset.