Evaluation of CNN Models with Transfer Learning in Art Media Classification in Terms of Accuracy and Class Relationship

Abstract

The accuracy obtained in Art MediaClassification (AMC) using CNN is lower comparedto other image classification problems, where theacceptable accuracy ranges from 90 to 99%. Thisarticle presents an analysis of the performance ofthree different CNNs with transfer learning for AMC, toanswer the question of what challenges arise in thisapplication. We proposed the Art Media Dataset (ArtMD)to train three CNNs. ArtMD contains five classes ofart: Drawing, Engraving, Iconography, Painting, andSculpture. The analysis of the results demonstrates thatall the tested CNNs exhibit similar behavior. Drawing,Engraving, and Painting had the highest relationship,showing a strong relationship between Drawing andEngraving. We implemented two more experiments,removing first Drawing and then Engraving. Thebest performance with 86% accuracy was achieved byremoving Drawing. Analysis of the confusion matrixof the three experiments for each CNN confirms thatDrawing and Painting have the lowest accuracy, showinga strong misclassification with the other classes. Thisanalysis presents the degree of relationship between thethree CNN models and details the challenges of AMC.