TIDAL vR.P.19_U

Tampered image detection and localization model built with a 9-channel multi-quality RGB ELA input and a UNet with a ResNet-34 encoder.

Included files

• checkpoints/best_model.pt
• checkpoints/vR.P.19_U_unet_resnet34_mqela_rgb.pth
• notebooks/vR.P.19_U Image Detection and Localisation.ipynb

Notes

• Model version: vR.P.19_U
• Primary task: image tampering localization with image-level tamper detection
• Input representation: 9-channel multi-quality RGB ELA at Q=75/85/95

Minimal load example

import segmentation_models_pytorch as smp
import torch

model = smp.Unet(
    encoder_name="resnet34",
    encoder_weights=None,
    in_channels=9,
    classes=1,
    activation=None,
)

state = torch.load("checkpoints/best_model.pt", map_location="cpu", weights_only=False)
weights = state.get("model_state_dict", state) if isinstance(state, dict) else state
model.load_state_dict(weights, strict=False)
model.eval()