Modified U-Net With VGG11-BN Encoder and Frost Filter Based Pre-Processing for Copy Move Forgery Localization

Abstract

Copy-move forgery is a widely used forgery method in digital image forensics, which involves copying and pasting a part of an image into another part of the image to hide or duplicate the visual information. This makes it difficult to detect such manipulations, particularly when the regions are subjected to geometric and photometric transformations. A Frost Filter-Enhanced Encoder-Decoder U-Net is proposed in this work to localize the copy-move forgery at the pixel-level with robustness. The proposed method combines these two technologies, namely adaptive Frost filtering and light-weight encoder-decoder segmentation architecture, to enhance noise suppression, structural feature preservation and forged boundary refinement. The encoder is a VGG11-BN (Batch Normalization) network to capture multi-level contextual information and the decoder is a skip-connected network to generate high-resolution forgery masks for precise localization of the manipulated region. The proposed framework is tested in the benchmark dataset CoMoFoD small v2 which includes four forgery transformations: Rotation, scaling, compression, blurring and noise addition. The experimental results showed that the proposed model obtained a Dice coefficient of 0.7977, Intersection over Union (IoU) of 0.7447, Precision of 0.7698, Recall of 0.8470, F1-score of 0.7977 and Accuracy of 0.9859. Moreover, Matthews Correlation Coefficient (MCC) and Cohen's Kappa of 0.7974 and 0.7927 were obtained, respectively, demonstrating good consistency in segmentation and excellent forgery discrimination ability.