Degradation Classification on Ancient Document Image Based on Deep Neural Networks

In this paper, we study degradation classification on ancient document images using three pre-trained models of benchmarking CNN architecture, i.e., Resnet101, Mobilenet V2, and Shufflenet. We use Document Image Binarization Contest (DIBCO), Persian Heritage Image Binarization Dataset (PHIBD), and private Jawi datasets for experimental purposes. We grouped the degradation into four categories, namely: bleedthrough/showthrough/ink-bleed, faint-text and low contrast, smear-spot-stain, and uniform degradation. In the training progress, we set optimizer to ADAM, the initial learn rate to 10
^-4
, and three epoch values: 5, 25, and 50 training epoch. To test the model, we conduct two testing stages: (1) unblind testing, (2) blind testing. The result shows that Shufflenet with 25 training epoch achieved 100% and 85% accuracy of unblind and blind testing, respectively, and obtained the fastest computational process. We concluded that Shufflenet could be chosen in classifying degradations based on its accuracy and computational time.