Bio-SFT: Asymmetric Cortical Guidance and Retinal Adaptation for Robust HDR Reconstruction

Spike-RetinexFormer: Rethinking Low-light Image Enhancement with Spiking Neural Networks

This work pioneers the synergistic integration of SNNs into Transformer architectures for LLIE, establishing a compelling pathway toward powerful, energy-efficient low-level vision on resource-constrained platforms.

BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement

Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input. Generative transport can model exposure ambiguity; however, its flexibility may also alter observable geometry and chromatic content. Moreover, fixed invertible color coordinates are usually treated only as representations, although their inverse mappings reshape the RGB-domain gradients received by the enhancement network. To address these issues, we propose Structure-Anchored Rectified Flow (SA-RF), which maintains correspondence through separate chromaticity/intensity stems, a scale-matched condition pyramid, and HybridAda. HybridAda assigns location-specific retrieval to spatial cross-attention and global exposure modulation to pooled AdaLN. We further introduce BC-IHV, a learnable Box--Cox polar color space whose analytically invertible intensity mapping controls the inverse-gradient dynamic range through a single exponent. This allows the representation to balance dark-range expansion and gradient conditioning instead of adopting a fixed linear or logarithmic law. Experiments on three LOL benchmarks, blind image-quality evaluation, and cross-dataset tests demonstrate consistent reconstruction and perceptual advantages over the sota. Controlled studies further support the effectiveness of both the proposed framework and color representation.

Keep It Frozen: Domain-Routed Conditional Residual Modulation for Multi-Domain Vision Transformers

OmniSEG: A Multi-Model Ensemble Framework for Retinal Layer Segmentation in OCT Images

Automated Accurate segmentation of retinal layers from optical coherence tomography (OCT) volumes is a prerequisite for the quantitative assessment of neurodegenerative and macular diseases. Yet the task is complicated by low interlayer contrast, imaging noise, and wide anatomical variability across patients. In this work we introduce OmniSEG, an ensemble framework that combines a nested UNet with dense skip connections (UNet++), a Swin Transformer-based UNet (SwinUNet), and an attention-guided surface auto-encoder into a unified training pipeline. The model is trained on the publicly available GOALS-2022 dataset using a compound loss that mixes focal cross-entropy, Tversky, gradient-based boundary, and surface smoothness terms while remaining numerically stable under automatic mixed precision (AMP). After 200 epochs the ensemble reaches a mean Dice of 0.864 across six foreground retinal classes, with a surface RMSE decreasing from 122 pixels to 76.5 pixels. Per-class analysis reveals that the OPL class benefits earliest from training (Dice > 0.10 at epoch 197), which reflects early-stage learning on a relatively wider layer rather than a final result, consistent with its wider span and higher contrast relative to thinner layers such as RNFL. These results establish a reproducible baseline for multi-class OCT segmentation and identify concrete directions for future improvement.

DSF-Net: dual selective fusion network with spatial-frequency domain encoding for retinal vessel segmentation

Retinal vessel segmentation is a fundamental task in quantitative fundus image analysis. However, existing methods still face challenges in segmenting thin and complex vessels because local details and global contextual information are often insufficiently integrated. To address this issue, we propose a Dual Selective Fusion Network (DSF-Net) for retinal vessel segmentation. The proposed network consists of a Dual-Branch Encoder (DB-Encoder), a Pinwheel-based Local Attention (PLA) module, and a Dual Selective Fusion Transformer Block (DSFTB). The DB-Encoder jointly models spatial- and frequency-domain information to capture both fine vessel details and global contextual patterns. The PLA module enhances local perception and boundary sensitivity through asymmetric multidirectional convolutions and Sobel edge priors. The DSFTB integrates Multi-scale Feature Attention (MSFA) and token-selective Global Feature Attention (GFA) to enable adaptive feature fusion and long-range dependency modeling. Experiments conducted on the DRIVE, STARE, and CHASE_DB1 datasets demonstrate that DSF-Net achieves competitive overall performance compared with existing methods. In particular, the proposed method produces more accurate and structurally coherent segmentation results, especially for thin and complex vessels. These findings indicate that the combined modeling of local detail, frequency-aware representation, and global contextual dependency is effective for retinal vessel segmentation. DSF-Net provides a robust framework for improving vessel continuity and boundary delineation in fundus images. The source code is available at: https://github.com/liang050629/DSF-Net .

SVRCL-SR: a high spatial resolution imaging method for large-size plate-shaped components

Sparse-view rotational-scanning computed laminography (SVRCL) is an essential technique for the rapid inspection of large-size plate-shaped components. However, its reconstructed results inevitably suffer from severe noise and streak artifacts, which degrade detection reliability. Recently, deep learning-based techniques such as projection domain constraints and image domain post-processing have shown promising application prospects. But two major bottlenecks remain. First, convolutional neural networks struggle to capture the long-range distribution characteristics of streak artifacts. Second, self-attention mechanisms introduce additional computational overhead. To solve the above problems, we devise a dual-domain joint image super-resolution (SR) framework (SVRCL-SR). First, shallow structural features are extracted through convolutional layers. Then, we introduce a feature denoising module that focuses on the mid- to high-frequency regions and dynamically performs dual-domain joint denoising. Finally, large-kernel attention is realized via frequency-domain convolution and element-wise multiplication, which compensates for missing high-frequency information while reducing computational cost. Comprehensive evaluations on multiple datasets and SR scales indicate that the SVRCL-SR achieves superior performance in artifact suppression and high-frequency detail restoration, along with strong robustness.