CogSENet: Blind Image Deblurring with Blur-Conditioned Semantic Routing and Explicit Frequency Fusion

---

CogSENet: Blind Image Deblurring with Blur-Conditioned Semantic Routing and Explicit Frequency Fusion

Authors: Pan Wang, Yihao Hu, Xiujin Liu
arXiv ID: 2606.30030 (ECCV 2026)

Problem: Blind image deblurring methods struggle with real-world spatially varying degradations and lack the semantic awareness needed to distinguish valid textures from artifacts.

Key Methodology:

• Semantic-Driven State Space Module (SDSSM) - replaces uniform 2D scan with differentiable Gumbel-Softmax token routing into semantic groups, then applies prompt-modulated selective scan for non-causal global dependency modeling at linear complexity.
• Bi-Frequency Fusion Block (BFFB) - uses Haar wavelet transforms to explicitly decouple features into high-frequency (texture) and low-frequency (structure) components, processed through specialized branches and adaptively fused via a learnable scalar.
• Continuous Blur Field (CBF) + CLIP Joint Modulation - estimates a dense 2D displacement field from the blurry input and fuses it with frozen CLIP semantic embeddings via cosine-similarity attention to modulate deepest bottleneck features.

Key Results:

• GoPro: 34.72 dB PSNR / 0.9744 SSIM with only 8.9M params (Our); Our+ reaches 34.91 / 0.9757 with 19.4M params.
• Outperforms FFTformer (34.21 dB, 16.6M params) and EVSSM (34.51 dB, 17.1M params) while using ~50% fewer parameters.
• RealBlur-R: 41.83 PSNR / 0.9799 SSIM; RealBlur-J: 34.54 / 0.9473.
• Rain100H: 32.15 PSNR (+0.69 dB over Restormer); SOTS dehazing: 28.72 PSNR (+0.51 dB over MPRNet).

Applied Context: This is a drop-in deblurring backbone that achieves SOTA with far fewer parameters, making it practical for deployment in camera pipelines, computational photography, or mobile imaging apps. The SDSSM and BFFB modules are architecture-agnostic and could be ported into other low-level vision tasks (deraining, dehazing, denoising) as demonstrated.