GatedFrequencyIMDN

Lightweight (≤0.5M params) image super-resolution (x2/x3/x4) for the AI1003 course project. We train from scratch on DF2K and evaluate on Set5 (PSNR on Y channel with 2-pixel shave) against Bicubic (baseline 32.25 dB).

What’s new (our innovation)

• DFT-aware channel attention (DFTAttention): uses rfft2 amplitude, radial low/mid/high frequency masks, and learnable band gates to reweight channels with negligible parameter overhead.
• Gated frequency IMDB (GateFreqIMDB): multi-distillation blocks fused with the above frequency attention to preserve high-frequency details while staying lightweight.
• Multi-scale upsampling: supports ×2/×3/×4 via pixel shuffle with shared lightweight backbone.
• Parameter budget friendly: default num_feat=48, num_blocks=6, distill=0.25 keeps total params <0.5M (verify with sum(p.numel() for p in model.parameters())).

Environment

conda create -n ai1003 python=3.8 -y
conda activate ai1003
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
python setup.py develop

Tested on 8×RTX 3090 (24GB) per course server.

Data

• Server path: /mnt/8thdd/AI1003/public (DF2K, Set5). Copy/symlink into datasets/ following datasets/README.md.
• Train: DF2K. Test: Set5. Scale: x2 (primary). Eval on Y channel, shave 2 px.

Key files

• basicsr/archs/GateFreqIMDN_arch.py — core model (DFTAttention + GateFreqIMDB + multi-scale upsampling).
• basicsr/train.py, basicsr/test.py — entrypoints.
• options/train/*.yml, options/Test/*.yml — configs (set dataset paths, scale, logging, seeds).
• experiments/ — logs/checkpoints. results/ — test outputs.

Training

• Set your train config (e.g., options/train/train_gated_freq_imdn.yml): data roots, scale, crop size, batch size, epochs, optimizer/LR schedule, losses, augmentations.
• Run:

python basicsr/train.py -opt options/train/train_gated_freq_imdn.yml

• Check param count (goal ≤0.5M) and log it.

Testing

• Set test config (e.g., options/Test/test_gated_freq_imdn.yml): dataset roots, checkpoint path, use_chop if memory-limited.
• Run:

python basicsr/test.py -opt options/Test/test_gated_freq_imdn.yml

• Report PSNR on Y channel with 2-pixel shave; compare to Bicubic 32.25 dB.

Evaluation & reporting (course requirements)

• Baseline: Bicubic 32.25 dB on Set5 (x2). Extra credit per +0.2 dB.
• Report: PSNR, param count (≤0.5M), optionally FLOPs; qualitative comparisons.
• Train from scratch (no pretrained weights).
• Deliverables: PPT, PDF (CVPR 2026 template, 4–8 pages), arch file, configs, (custom) model file, checkpoints (experiments/pretrained_models).

Repro tips

• Fix seeds in configs; log hardware (8×3090), epochs, batch size, LR schedule, losses, aug (crop size, flip/rotate).
• Keep checkpoints under experiments/ and outputs under results/.
• For stability, default band gates are zero-initialized to near-uniform softmax; adjust masks or gating only if you change input resolution regimes.

Acknowledgements

Built on BasicSR.