Due to loss of server access permissions, partial reference code is provided. The implementation of EdgeCAM is based on the trochcam library, and only minor modifications to Grad-CAM++ are required to complete the computation . The project is not an end-to-end structure, and the training steps are as follows:

1. Train a tampering detection binary classification model, saving weights for each epoch (which consumes significant storage space);
2. Calculate candidate SOP (already provided);
3. Compute EdgeCAM visualizations using models from different training stages;
4. Fuse EdgeCAM results to obtain coarse tampered region edges;
5. Use EdgeCAM edges to select the final tampered region mask from candidate SOP.
   For continuous assistance, contact yzhoulv@foxmail.com.

SOP：https://pan.baidu.com/s/1PXUHecO7TT66DUikOh4epg?pwd=vfms code：vfms

ps-battles dateset: https://drive.google.com/file/d/10E5Trc3r_ScE4VeyrZ0tGLKM7HTCu3xt/view?usp=drive_link