Research map / Label-efficient and robust learning / Recognition and long-tail
A Unified Framework for Microscopy Defocus Deblur with Multi-Pyramid Transformer and Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a multi-pyramid transformer (MPT) with cross-scale window attention and an extended frequency contrastive regularization (EFCR) for microscopy defocus deblurring, with a new surgical microscopy blur dataset.
Problem
Defocus blur in microscopy (cell and surgical) degrades images and downstream tasks; microscopy images need longer attention spans than natural images and suffer from scarce blur-sharp training data.
Approach
A U-shaped transformer whose pyramid blocks use cross-scale window attention (local queries vs. downscaled key-values) plus intra-scale channel attention fused by an asymmetric-activation FEFN, combined with a wavelet-domain contrastive loss that uses synthetic reblurring and cross-domain extra data (LFDOF) to exploit latent deblur signals.
Key results
- MPT+EFCRex reaches 35.16 dB PSNR on BBBC006w1 versus 34.79 for Restormer, and 26.27 dB on DPDD vs 25.98 for Restormer.
- On cell detection over BBBC006, MPT+EFCRex improves mean AP by 19.57% over blurry input, versus 16.05% for Restormer and 13.50% for GRL.
- MPT uses 76 FLOPs and 19.80M parameters versus Restormer's 141 FLOPs and 26.12M, with the paper reporting roughly 46% fewer FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MPT instead of the strongest listed baseline | microscopy deblurring · BBBC006w1 | PSNR: 34.79 -> 34.89 | Helped |
| use MPT instead of the strongest listed baseline | microscopy deblurring · BBBC006w2 | PSNR: 29.78 -> 29.85 | Helped |
| use MPT instead of the strongest listed baseline | microscopy deblurring · 3DHistech | PSNR: 33.49 -> 33.55 | Helped |
| use MPT instead of the strongest listed baseline | microscopy deblurring · CaDISBlur | PSNR: 44.86 -> 44.98 | Helped |
| add EFCR to MPT (ablation) | MPT · BBBC006w1 | PSNR: 34.89 -> 34.96 | Helped |
| add EFCR to MPT (ablation) | MPT · DPDD | PSNR: 26.21 -> 26.23 | Helped |
| add LFDOF as extra data with EFCRex (ablation) | MPT+EFCR · BBBC006w1 | PSNR: 34.96 -> 35.16 | Helped |
| add LFDOF as extra data with EFCRex (ablation) | MPT+EFCR · DPDD | PSNR: 26.23 -> 26.27 | Helped |
| replace CSWA with window attention (ablation) | MPT · DPDD, BBBC006w1, CaDISBlur | PSNR: 26.21 -> 26.13; 34.89 -> 34.10; 44.98 -> 44.78 | Hurt |
| replace CSWA and ISCA with window attention (ablation) | MPT · DPDD, BBBC006w1, CaDISBlur | PSNR: 26.21 -> 25.92; 34.89 -> 33.98; 44.98 -> 44.02 | Hurt |
| replace the FEFN asymmetric activation with concatenation followed by GELU (ablation) | MPT · DPDD, BBBC006w1, CaDISBlur | PSNR: 26.21 -> 26.18; 34.89 -> 34.86; 44.98 -> 44.84 | Hurt |
| replace the FEFN asymmetric activation with addition followed by GELU (ablation) | MPT · DPDD, BBBC006w1, CaDISBlur | PSNR: 26.21 -> 26.03; 34.89 -> 34.75; 44.98 -> 44.87 | Hurt |
Surprising findings
- Naive pretraining on extra data then fine-tuning gave trivial or negative gains (e.g., -0.02 dB on BBBC006w1), while EFCRex gave larger gains.
Limitations
- Evaluation on larger-scale extra datasets such as ImageNet is left for future work.
- Ablations and dataset details are deferred to supplementary material.
Experimental setup
Datasets: DPDD LFDOF BBBC006 3DHistech CaDISBlur CUHK WNLO CataBlur CaDIS
Benchmarks: DPDD BBBC006 3DHistech CaDISBlur WNLO CataBlur CUHK
Metrics: PSNR SSIM LPIPS AP mIoU PA FLOPs
Baselines: DRBNet GKMNet MIMO-UNet MSSNet SwinIR PANet GRL Restormer KPAC IFAN NRKNet
Builds on: Restormer GRL DRBNet contrastive regularization
Closest papers by meaning
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.85)
- ChAda-ViT : Channel Adaptive Attention for Joint Representation Learning of Heterogeneous Microscopy Images (CVPR 2024 · similarity 0.84)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.83)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.83)
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.83)
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.