The high dimensionality and inherent noise in multi-omics data complicate analysis, making multi-view feature selection a vital solution for integrating heterogeneous data. However, existing methods still suffer from the following limitations: (i) They neglect structural information among features within individual views; (ii) They apply uniform penalties to all features, failing to distinguish between the importance of feature groups and independent features. This one-size-fits-all strategy may over-penalize cohesive feature groups containing useful information or retain less informative features. Furthermore, direct computation of pairwise fuzzy mutual information is expensive in high dimensions. To this end, this paper proposes a Fuzzy-weighted Group-aware Unsupervised Multi-view Feature Selection (FWGAFS) method. This method adaptively penalizes features and feature groups according to their importance, with penalty strength inversely proportional to their evaluated significance. The resulting regularization is intended to favor informative features and reduce the influence of less informative ones. Specifically, FWGAFS first employs density-based clustering to partition features into groups, capturing view-specific omics structures. It then uses an anchor-based strategy to reduce the computational overhead of fuzzy mutual information. Subsequently, a fuzzy-weighted group-aware regularization term is formulated to implement differentiated penalties. Concurrently, FWGAFS integrates unsupervised clustering with graph learning to capture underlying data structures, thereby guiding the selection of structurally relevant features. An alternating optimization algorithm solves the proposed model. Experiments on real-world multi-omics datasets show that FWGAFS achieves competitive overall performance compared with the state-of-the-art methods.
Single-modal visual positioning systems face critical limitations in complex environments, particularly in operational coverage and robustness. This paper presents a hybrid positioning system to overcome these issues. It combines a Fisheye Image Sensor (FIS), a Pinhole Image Sensor (PIS), and a photodetector (PD), which jointly improves positioning coverage and accuracy. The system employs a quadratic LED Array as the positioning beacon, with the PD performing rough positioning and a heterogeneous binocular vision architecture enabling precise positioning. Through the establishment and analysis of an FIS error model, a depth-compensated vision positioning algorithm is proposed to mitigate radial distortion. Subsequently, an error arbitration-based heterogeneous vision positioning method is designed by comparing the range models of PIS and FIS, enabling dynamic switching between the two sensors via error-optimized modeling. Experimental results demonstrate: (1) Under illumination from a single quadratic LED array indoors, the system achieves a 4× expansion in positioning range, covering a 6m-diameter area at 3.0m height. (2) In a 6×6×3 m3 volume, the system achieves an average positioning error of 8.32 cm in near-field zones and 12.87 cm in far-field zones, yielding a total mean error of 9.712 cm.
Portrait cast shadow removal is a critical yet challenging task in computer vision and portrait editing. Current methods mainly rely on planar features, such as facial landmarks, to assist in cast shadow removal and often fail to fully exploit the geometric information of the face. Moreover, current methods often ignore the asymmetry of facial details and self-shadows caused by lighting orientations. In this work, we propose a Portrait Shadow Diffusion (PSDiffusion) framework to address these challenges. PSDiffusion combines diffusion models with facial geometry and illumination, achieving delicate balance between cast shadow removal and facial self-shadow preserving through illumination consistency constraint and symmetry-aware attention mechanism. Specifically, we first design a Face Symmetry Flip Module to generate a flipped facial image, which presents the potentially cast shadow-free features for subsequent shadow removal. Then, we introduce a pre-trained Feature Extraction Diffusion Module to extract cast shadow free features from the flipped facial image. Finally, we adaptively fuse the extracted cast shadow free features with the features from the input image into the Symmetry-aware Diffusion Module, and generate the final cast shadow removal result with preserved self-shadows. Furthermore, we construct a large-scale portrait shadow dataset (LPSD) with diverse real-world scenarios of portrait shadows. Comprehensive experiments conducted on existing open-source facial datasets and web-collected images demonstrate that our method surpasses state-of-the-art techniques in cast shadow removal.
Deep unfolding networks (DUNs) have significantly advanced image restoration. However, most unfolded architectures pass information strictly stage by stage, assuming progressive quality improvement. This overlooks intermediate reconstructions as historical candidates whose local reliability varies spatially. We observe that restoration quality does not strictly increase across iterations; earlier stages often preserve fine textures and subtle structures lost in the final output. Based on this insight, we propose MTADUN, a task-adaptable state-transmission unfolding framework for image restoration, where intermediate reconstruction states are explicitly preserved, structurally regularized, and selectively reused along the unfolding trajectory across diverse degradation settings. Specifically, MTADUN maintains intermediate image-domain reconstructions as a historical state pool and performs current-conditioned local retrieval over these candidate states. A wavelet tree-structured prior encourages cross-scale structural coherence, a multi-stage prior regularizes abrupt trajectory variations, and a cross-stage transmission-augmented mechanism adaptively fuses locally compatible historical states. We use compressed sensing as the primary diagnostic setting to analyze the non-monotonic local reliability of intermediate states, and further instantiate the same state-transmission principle on denoising, deraining, and deblurring with task-specific training. Extensive experiments, diagnostic analyses, and modular replacement ablations demonstrate the effectiveness and broad applicability of the proposed framework.
Semi-supervised medical image segmentation aims to reduce the burden of manual annotation by leveraging a limited set of labeled data along with a large collection of unlabeled images, while maintaining reliable segmentation performance for clinical use. However, most existing methods neglect the consistency of boundary predictions across different learning views, often resulting in imprecise segmentation along object edges. In this paper, we propose a novel Collaborative Boundary enhancement with Difficulty Awareness (CoBDA) framework for semi-supervised medical image segmentation. Specifically, we propose a Consistency-aware Boundary Enhancement (CBE) module, which utilizes the agreement between predictions from two subnets to reinforce reliable edge signals and suppress ambiguous boundaries. Furthermore, to improve semantic coherence across the dual networks, we propose a Cross-prototype Semantic Alignment (CSA) strategy that facilitates mutual semantic representation alignment, thereby encouraging prediction consistency. Finally, we propose a Difficulty-aware Dynamic Weighting (DDW) strategy, which adaptively focuses the model’s attention on challenging regions. By integrating boundary cues and uncertainty, DDW ensures the learning process prioritizes hard-to-segment pixels. Extensive experiments on four benchmark datasets demonstrate that our model significantly outperforms state-of-the-art semi-supervised medical image segmentation methods. The code will be released upon acceptance at https://github.com/taozh2017/CoBDA.
