MSDCT-UNet

Multi-Scale Discrete Cosine Transform UNet for local motion blur detection.

Checkpoint

• bocchi.pth — trained on the BOCCHI dataset.

Usage

import torch
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="BrianChen54923/msdct-unet",
    filename="bocchi.pth",
)
state_dict = torch.load(ckpt_path, map_location="cpu")

# Load into your MSDCT-UNet model definition
# model.load_state_dict(state_dict)

Input: RGB image at 1080 × 720. Output: single-channel blur probability mask (0 = sharp, 1 = blur).

License

CC BY-NC 4.0 — research use only.