Chaolong Zhang

Memory-bank-based 3D point cloud anomaly detection identifies defects via nearest-neighbor retrieval of nominal features. However, current methods often neglect surface normals and suffer from inadequate coverage of complex geometries due to random coreset initialization. We propose a unified framework to address these gaps via two integrated components: Normal-Aware Registration Features (NARF), which augments the RANSAC--ICP pipeline to extract per-point orientations with negligible overhead, and Geometry-Aware Coreset Sampling (GACS), which prioritizes high-curvature regions using curvature- and normal-variation-weighted sampling. By fusing coordinate, semantic (PointMAE), and angular-deviation streams through rank-normalization, our approach achieves 84.4% O-AUROC on the Real3D-AD benchmark, delivering state-of-the-art performance with enhanced geometric discriminability.

Micro-expression recognition (MER) is a fundamental yet challenging task in affective computing due to the subtle, transient, and localized nature of spontaneous facial muscle movements. Although Vision Transformers effectively extract MER features, spatial regions are often processed in isolation, limiting the ability to capture long-range relational dependencies across facial zones. To address this problem, this study proposes a so-called HTNet-SDGCN that augments a hierarchical Transformer backbone with a Structural-Dynamic Dual Graph Convolutional Network (SD-GCN). SD-GCN models inter-region dependencies via two complementary graphs, i.e., a shared structural graph that learns stable facial topology across all training samples, and a dynamic graph that adapts to the motion pattern of each individual input. These graphs are adaptively fused to generate a comprehensive relational representation. Furthermore, Landmark-Guided Gaussian Weighting is introduced to sharpen optical-flow signals around semantic facial landmarks. Extensive experiments on the CASME II, SMIC, and SAMM datasets under the leave-one-subject-out (LOSO) protocol demonstrate that the proposed framework achieves robust performance, attaining a UF1 of 0.8813 and a UAR of 0.8661 on the full composite dataset, and reaching a peak UF1 of 0.9564 on CASME II.

Hyperspectral imagery (HSI), with its rich spectral information across continuous wavelength bands, has become indispensable for fine-grained land cover classification in remote sensing applications. Although some existing deep neural networks have exploited the rich spectral information contained in HSIs for land cover classification by designing some adaptive learning modules, these modules were usually designed as additional submodules rather than basic structural units for building backbones, and they failed to adaptively model the spectral correlations between adjacent spectral bands and nonadjacent bands from a local and global perspective. To address these issues, a new adaptive spectral-correlation learning neural network (ASLNN) is proposed for HSI classification. Taking advantage of the group convolutional and ConvLSTM3D layers, a new adaptive spectral correlation learning block (ASBlock) is designed as a basic network unit to construct the backbone of a spatial–spectral feature extraction model for learning the spectral information, extracting the spectral-enhanced deep spatial–spectral features. Then, a 3D Gabor filter is utilized to extract heterogeneous spatial–spectral features, and a simple but effective gated asymmetric fusion block (GAFBlock) is further built to align and integrate these two heterogeneous features, thereby achieving competitive classification performance for HSIs. Experimental results from four common hyperspectral data sets validate the effectiveness of the proposed method. Specifically, when 10, 10, 10 and 25 samples from each class are selected for training, ASLNN achieves the highest overall accuracy (OA) of 81.12%, 85.88%, 80.62%, and 97.97% on the four data sets, outperforming other methods with increases of more than 1.70%, 3.21%, 3.78%, and 2.70% in OA, respectively.

Surface defect detection is crucial in industrial production, and due to the conveyor speed, real-time detection requires 30–60 frames per second (FPS), which exceeds the capability of most existing methods. This demand for high FPS has driven the need for lightweight detection models. Despite significant advancements in deep learning-based detection that have enabled single-stage models such as the you only look once (YOLO) series to achieve relatively fast detection, existing methods still face challenges in detecting multi-scale defects and tiny defects on complex surfaces while maintaining detection speed. This study proposes a lightweight single-stage detection model called lightweight industrial defect detection network with improved YOLO architecture (LIDD-YOLO) for high-precision and real-time industrial defect detection. Firstly, we propose the large separable kernel spatial pyramid pooling (SPP) module, which is a SPP structure with a separable large kernel attention mechanism, significantly improving the detection rate of multi-scale defects and enhancing the detection rate of small target defects. Secondly, we improved the Backbone and Neck structure of YOLOv8n with dual convolutional (Dual Conv) kernel convolution and enhanced the faster implementation of Cross Stage Partial Bottleneck with 2 Convolutions (C2f) module in the Neck structure with ghost convolution and decoupled fully connected (DFC) attention, reducing the computational and parameter overhead of the model while ensuring detection accuracy. Experimental results on the NEU-DET steel defect datasets and printed circuit board (PCB) defect datasets demonstrate that compared to YOLOv8n, LIDD-YOLO improves the recognition rate of multi-scale defects and small target defects while meeting lightweight requirements. LIDD-YOLO achieves a 3.2% increase in mean average precision (mAP) on the NEU-DET steel defect dataset, reaching 79.5%, and a 2.6% increase in mAP on the small target PCB defect dataset, reaching 93.3%. Moreover, it reduces the parameter count by 20.0% and floating point operations by 15.5%, further meeting the requirements for lightweight and high-precision industrial defect detection models.

Fringe projection profilometry (FPP) is a widely used non-contact 3D measurement method. Though maturing in the last decade, it remains a significant challenge when facing the phase unwrapping of measured object surfaces in a single-shot measurement setting. With the rapid development of deep learning techniques, the adoption of a data-driven approach is gaining popularity in the field of optical metrology. This study proposes a new absolute phase recovery method based on the devised single-stage deep learning network. The aim is to ensure high-quality absolute phase recovery from a single-shot fringe projection measurement. Unlike most existing approaches, where the numerators and denominators of the wrapped phases and the fringe orders are predicted in various stages, the proposed method acquires the wrapped phases and the corresponding fringe orders within a single network, i.e. it can predict both wrapped phases and the corresponding fringe orders directly and simultaneously from the single fringe pattern projected in the single-shot mode based on a unified Y-shaped network. Experiments on benchmark datasets and models have demonstrated the effectiveness and efficiency of the technique, especially in terms of high-quality recovery of absolute phase information by using the lightweight single-stage network, and enabling the FPP-based phase 3D measurements in an online manner.

With the development of intelligent manufacturing, the production and assembly accuracy of components in factories is increasing in line with growing demand. However, the traditional manual quality inspection is inefficient, inaccurate, and costly. To this end, digital and optical imaging techniques are used to achieve intelligent quality inspection. However, during the reconstruction process, the high reflectivity of object materials affects the speed and accuracy of reconstruction results. To overcome these problems, this study investigated the three-dimensional (3D) digital imaging techniques based on line laser scanning. It advances a novel methodology for image segmentation, underpinned by deep learning algorithms, to augment the precision of the reconstruction results while simultaneously enhancing processing velocity. After the reconstruction phase, the research assesses flatness tolerance using point cloud registration technology. Finally, we constructed a measurement platform with a cost of less than CNY 100,000 (about USD 14,000) and obtained a measurement accuracy of 30 microns.