Publication: Image inpainting detection using deep learning network
Date
2021-07-01
Authors
Tew, Jin Chun
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
An approach of image inpainting detection with Convolutional Neural Network (CNN) is proposed in this research. Image inpainting is a method in image forgery and is often used by unethical users such as removing watermarks. There were some researches on detecting image inpainting but most of them are using conventional approach. There are limitations on these traditional methods such as low accuracy and need to select suspicious region manually in prior. Thus, transfer learning of ResNet50 is proposed for image inpainting classification. U-Net with ResNet50 backbone is the proposed method for localizing the inpainted region. U-Net is the encoder-decoder network that predict the mask whereas ResNet50 serves as the feature extractor in the encoder path. The classification model is trained and tested with 100 real images and 100 patch-based inpainted images. After k-fold cross validation, the average classification accuracy is 83.0%. The localization model is trained and tested with 100 patch-based inpainted images with its corresponding masks, the average F1_score after k-fold cross validation is 93.90%. It is proved that transfer learning of ResNet50 is able to classify patch-based inpainted images and U-Net with ResNet50
is able to localize the inpainted region.