Pei Zhou

Personal protective equipment (PPE) detection in industrial scenes is a joint classification-and-localization problem complicated by partial occlusion, scale variation, and background clutter. PPE categories also follow structured physical relations; for example, helmets are normally associated with heads and gloves with hands. Existing attention modules mainly reweight appearance features and seldom encode these directional inter-class dependencies. We therefore propose IRA-YOLO, a YOLO-based object detector equipped with an Inter-Class Relation Attention Module (IRAM). The IRAM projects each multi-scale feature map into class-specific subspaces, learns asymmetric pairwise relations among category attention maps, and returns the relational prior to the detection feature through residual fusion. On SH17, IRA-YOLO improves mAP@50:95 from 0.396 to 0.411 over YOLOv11s while adding 0.42 M parameters and 0.99 GFLOPs, and it maintains 139.83 FPS in the single-image inference test. Cross-dataset evaluation on CHVG further supports the transferability of the proposed feature refinement. These results indicate that explicit inter-class relation modeling provides complementary semantic context for multi-class PPE detection.

Fringe projection profilometry is widely used for three-dimensional surface measurement, but high dynamic range (HDR) effects can degrade measurement accuracy. Conventional multi-exposure methods improve reconstruction quality at the cost of complex and time-consuming data acquisition, while many deep learning-based methods rely on a single overexposed fringe image and may therefore perform poorly under severe HDR conditions. To address this limitation, we propose a fusion-guided restoration network (FGNet) that learns pixel-wise modulation-aware reliability maps from high- and low-exposure fringe images. A modulation-complementary fusion module exploits the complementary information in the two inputs, and a guided restoration module further refines the fused result. Experiments demonstrate that FGNet effectively restores HDR fringe patterns and improves reconstruction accuracy over existing methods. For the metal cover plate, it achieves a reconstruction MAE of 0.0113 mm.

Existing self-supervised deep learning methods for single-pixel imaging (SPI) often struggle to simultaneously achieve high reconstruction quality and computational efficiency. To address this limitation, this paper proposes FDE-Net, an efficient self-supervised SPI algorithm based on frequency-domain enhancement. FDE-Net adopts a single-path encoder–decoder architecture and reduces computational complexity by implementing feature upsampling through a lightweight efficient up-convolution module. To ensure high-fidelity target reconstruction, a frequency-domain brightness adaptive correction module is introduced alongside a multi-scale attention fusion strategy to enhance the network’s feature representation capability. Experimental results on the Set11 dataset demonstrate that FDE-Net achieves the best reconstruction quality among all compared methods. At a 25% sampling rate, it yields an average PSNR improvement of 2.1 dB over the algorithm SDD-Net while reducing reconstruction time by approximately 33.3%, effectively balancing reconstruction quality and computational efficiency.

Personalized oncology treatment recommendation is a critical clinical task that requires in-tegrating complex, multi-modal patient data with established medical knowledge to ensure both accuracy and safety. While deep learning models excel at capturing latent patterns from high-dimensional data, their opaque decision-making processes and inability to strictly enforce clinical constraints hinder their adoption in high-stakes medical domains. Conversely, traditional rule-based systems offer high interpretability but struggle to scale with complex, heterogeneous data. To address these challenges, we propose the Knowledge-driven Neuro-Symbolic Network (K-NeSyNet), a novel framework for personalized oncology treatment recommendation. K-NeSyNet is grounded in a newly constructed Multi-Modal Oncology Knowledge Graph (MM-OKG) that unifies genomic mutations, medical imaging features, clinical text, and structured medical guidelines from publicly available sources including TCGA, DGIdb, KEGG, and NCCN guidelines. The core innovation of K-NeSyNet is a three-channel differentiable symbolic reasoning mechanism that explicitly models guideline recommendations, mutation-target matching, and contraindication penalties for each patient-drug pair. These symbolic signals are dynamically fused with the outputs of a knowledge-aware graph attention neural reasoning module via an adaptive gated fusion network. Crucially, the fusion gate learns to balance neural and symbolic confidence in a patient-specific manner, while contraindication evidence enters both the symbolic score and a dedicated safety objective to down-weight clinically risky drugs. Extensive experiments on a real-world multi-modal oncology dataset comprising 4,781 patients across 10 cancer types demonstrate that K-NeSyNet consistently outperforms eight state-of-the-art baselines. Specifically, K-NeSyNet achieves the highest F1@10 of 0.9227, NDCG@10 of 0.9656, and Jaccard similarity of 0.9366, while maintaining competitive Clinical Guideline Consistency. Ablation studies confirm the indispensable role of each component, with the removal of the fusion gate causing the most significant performance degradation. Furthermore, K-NeSyNet provides transparent, score-decomposed explanations for its recommendations, offering a crucial step toward trustworthy AI-assisted clinical decision support.