A multi-task image splicing tampering detection model based on semantic segmentation network

Abstract

Abstract Nowadays, malicious tampering and forgery images show an explosive growth trend. Current image splicing tamper detection methods commonly suffer from issues such as inaccurate identification of tampering boundaries and low detection accuracy. To solve these problems, this paper proposes a multi-task image splicing tampering detection algorithm based on DeepLabv3+. The proposed algorithm improves the model's detection of manipulation boundaries by using shallow image features to predict the boundaries of manipulation regions within images. Concurrently, in an endeavor to enhance the model's adaptability towards multi-scale manipulated regions, segments the image manipulated regions by multi-scale fusion features, integrates the attention mechanism in the original empty-space pyramidal kernel, and uses the MSPM instead of the global average pool to improve the localization accuracy of the manipulated regions. The experimental results demonstrate that the F1 scores on CASIAV1.0 and Columbia datasets are 0.664 and 0.729, respectively. The MCC scores are 0.683 and 0.723, respectively. These scores are superior to the algorithms CNN-LSTM, MFCN, EXIF-SC, and MSGL. Furthermore, Significant advancements have been observed in terms of the localization accuracy when comparing against the original methodologies.