Haochen Qi

Event factuality prediction, a task in natural language processing, estimates the commitment conveyed by a text toward an event. Nested negation and modality require inference sensitive to the linguistic scope, but endpoint labels can leave intermediate transformations underdetermined. We introduce SCOPE-Fact, which learns five-state stochastic operators from isolated lexical and grammatical signatures and composes them in scope order. Parent–child commitment judgments constrain local transformations, while a log-determinant acquisition criterion allocates a fixed annotation budget. Unsupported inputs use an independent text classifier shared across mechanistic comparisons. Our analysis establishes conditional recovery and perturbation bounds, gives a necessary total-variation diagnostic for local representation adequacy, and characterizes observable order effects through commutators. With five ratings per proposition, the simultaneous adequacy bound is inconclusive at the population level. The reported empirical scores are 51.5% macro-F1 on MAVEN-FACT public development and 69.0%, 59.0%, and 48.0% pair accuracy under combination, order, and depth holdouts. Relative to an equally supervised nonlinear recurrence, the point differences are 0.5, 4, 4, and 5 percentage points. The combination-holdout difference from the no-anchor ablation is 11 points. These descriptive comparisons are interpreted jointly with routing coverage and paired uncertainty. The framework makes the predictive reuse of local commitment transformations testable within a defined linguistic core, while the mathematical analysis separates prediction from population identifiability and representation adequacy.

Aero-engine defects often exhibit micro-scale and high-frequency characteristics under complex metallic textures, which makes precise segmentation difficult. Most existing pixel-level methods rely on spatial-domain modeling and lack frequency-domain decoupling. As a result, high-frequency details are easily hidden by low-frequency background information. In addition, repeated downsampling weakens the representation of fine-grained structures, leading to inaccurate boundary localization and limited robustness. To address these issues, a spectrum-driven hierarchical learning network is proposed for aero-engine defect segmentation. First, a dual-band spectral module is constructed using the discrete cosine transform to separate high-frequency and low-frequency components, providing stable and physically meaningful frequency-domain priors for the network. Second, a detail-guided module is designed where high-frequency features adaptively guide skip connections, compensating information loss during encoding and improving boundary recovery. Furthermore, a low-frequency-driven region-aware modeling module is developed. The internal defect regions, boundary areas, and background regions are modeled hierarchically. A dynamic hyper-kernel generation mechanism performs region-sensitive convolutional modeling, improving adaptation to complex structural variations. Extensive experiments on the Turbo19 and NEU-Seg datasets demonstrate that the proposed method produces accurate defect boundaries and achieves mIoU scores of 89.82% and 91.44%, improving over the second-best method by 5.22% and 4.42%, respectively.

In aeroengine maintenance, endoscopic imaging serves as a crucial tool for detecting blade defects and evolves toward intelligence driven by computer vision technology. Currently, supervised-learning-based defect segmentation methods mainly rely on extensive pixel-level annotations, making it laborious and time consuming. This article shifts focus to the abundant unlabeled data in real-world scenarios and introduces an innovative semisupervised defect segmentation method termed SAIT. Within this framework, three parallel self-supervised mechanisms are adeptly integrated with a semisupervised framework, aiming to bolster defect semantic segmentation with limited labeled samples. In the initial phase, by leveraging the capability of the vision transformer to dissect images into patches, four stochastic distortions are seamlessly infused into the patch sequence. Subsequently, three self-supervised tasks from image level to pixel level are achieved through a customized joint objective function paired with a tailored backbone network. In the second phase, SAIT undergoes pixel-level fine-tuning via the proposed class-centric loss, mitigating class imbalances in limited sample sizes and enhancing initial training. Experiments on a proprietary dataset demonstrate that SAIT achieved 78.42% and 86.70% in mean intersection over union and mean pixel accuracy metrics, respectively, with 25% labeled data, significantly improving the performance of existing semisupervised defect segmentation techniques. Meanwhile, experiments on the open-source dataset further indicate that SAIT holds promise for application in other industrial sectors beyond aeroengine inspection.

This study addresses the fine-grained segmentation of defects in bridge pavements, which is crucial for the maintenance and structural safety of bridges. Although bridge pavements pose distinctive challenges owing to their unique characteristics and varied defect types, previous studies have primarily focused on the detection of slender cracks. To fill this research gap, we developed a novel end-to-end hybrid method that dynamically combines the vision transformer (ViT) and level set theory to handle the complex geometry of bridge pavement defects. The novelty of the proposed method lies in the configuration of two parallel decoders. These decoders, operating under a unified objective function, share weights and perform simultaneous optimization, thereby facilitating a holistic end-to-end training process. Furthermore, we compiled two new bridge pavement defect datasets, namely BdridgeDefX and BdridgeDef20, which offer broader applicability for practical defect detection. The results of a rigorous experimental validation on four datasets demonstrated the proposed method’s capability of generating accurate defect boundaries and delivering state-of-the-art performance.

In remote sensing image analyses, the extraction of lake water bodies has been emphasized owing to its pivotal role in interpreting aquatic ecosystems, assessing hydrological trends, and detecting environmental changes. Although deep-learning-based techniques have been effectively deployed for this task, several challenges persist, including mis-segmentation of low-contrast regions, insufficient delineation of fuzzy boundaries, and over-segmentation in micro-regions. To address these issues, an innovative, end-to-end segmentation framework is proposed in this study. This framework ingeniously integrates the Geodesic Active Contour (GAC) model with the Vision Transformer (ViT) architecture, thereby offering a robust solution for fain-gained lake water body extraction. Specifically, the conventional GAC model is first revisited and reformulated. Subsequently, the Differentiable Spectral Clustering (DSC) module is designed for automatic contour initialization and lifts the dependence on hard threshold binarization. Finally, an iterative learning strategy is introduced to integrate the aforementioned models within the ViT architecture. This strategy serves to constrain the geometric characteristics, while also permitting dynamic adjustment of evolution parameters, facilitating enhanced refinement of spatial details and accuracy in segmentation results. Comprehensive experiments conducted on the newly constructed Global Lake View (GLV) dataset and a public benchmark demonstrated that the proposed method yields precise lake boundaries and achieves state-of-the-art performance.

With continuous advancements in sensor technology and computer vision, automated surface defect detection has become an important problem in the modern manufacturing industry. Although deep-learning-based defect detection methods have proven successful in various inspection tasks, they often require extensive high-quality pixel-level annotations for model training. Industrial defects, characterized by diverse shapes and intricate textures, require manual labelling by skilled inspectors. This study introduces a weakly supervised defect segmentation framework called the progressively learning dynamic level set (PLDL) to overcome these constraints. This framework incorporates two parallel learning modules and a differentiable level set module interconnected through a progressive learning strategy facilitated by an innovative loss function. Using only image-level labels as inputs, the PLDL framework iteratively refines object boundaries, dynamically optimizes the training process, and generates pixel-level outputs. Our method is efficient and consistent, requiring no manual supervision or post-processing steps. Experimental results on four benchmark datasets representing diverse industrial scenarios reveal that PLDL outperforms recent weakly supervised models. It achieved 76.52%, 85.23%, 85.71%, 84.43%, and 83.17% in mIoU, mPA, Precision, Recall, and F-measure, respectively, indicating its superior effectiveness in weakly supervised defect segmentation.

Deep supervised learning-based fault diagnosis methods require a large amount of labeled data, which frequently contradicts the typical engineering scenario, in which numerous samples are available but only a small portion are labeled. To conduct a fault diagnosis in this case, unsupervised and semi-supervised representation learning have recently been proposed. However, most of these methods have been designed without much consideration for downstream classification tasks and thus are insufficiently relevant for providing sufficient targeted assistance. Therefore, this study proposes a semi-self-supervised method that consists of two learning stages. In the self-supervised learning stage, two pretext tasks are implemented. Using two auxiliary feature extractors and classifiers, the global and local features of the unlabeled samples can be learned by the feature extractor. In the supervised learning stage, the feature extractor parameters are frozen, and the classifier is trained with the limited number of labeled samples available. The performance of the proposed method was verified using the CWRU open dataset and a self-built experimental dataset. Experiments on these two datasets demonstrated that the proposed method outperformed existing diagnosis methods for a few labeled samples.

With the rapid development of the aviation industry, videoscope inspection of aeroengines has become crucial for ensuring aircraft flight safety. Recently, deep learning, particularly convolutional neural networks (CNNs), have shown remarkable efficacy in videoscope inspection tasks. However, these methods usually require large-scale training labels with accurate annotation. In videoscope images, defects are often present at the micrometer level and require manual labeling by professional inspection personnel, posing further challenges to model training. To address these issues, weakly supervised deep level set (WDLS), a fully automatic framework for aeroengine defect segmentation, was proposed in this work. WDLS employs a multibranch structure and an iterative learning strategy to combine a high-performance CNN architecture with level set evolution. First, the similarity-guided region detector was designed to generate class-specific segmentation proposals and initialize the level set function. Second, the adaptive local parameter prediction algorithm was proposed to integrate local priors and constraints into the energy function and optimize the iterative process. Finally, a self-supervised objective function named convexified level set was introduced to represent an improved level set formulation and obtain the final segmentation. The proposed framework is continuously differentiable and unified. Thus, it can be seamlessly embedded in any CNN without postprocessing. Furthermore, the method was evaluated on a new aeroengine dataset Turbo19 and a benchmark dataset. Experimental results demonstrated that the proposed framework meets real-time requirements and achieves better performance than state-of-the-art methods.

X-ray nondestructive testing means are widely used in the inspection process of internal defects of parts. In practical inspection, defects are generally determined and rated by manual inspection based on X-ray images, which is inefficient and cannot meet the requirements of high-volume automatic inspection. This paper proposes an automatic defect classification model based on an improved U-Net. First, a classifier is added behind the encoder of U-Net. The encoder and classifier are connected in series to form the main branch of the model to complete the classification task. Second, the decoder of U-Net is improved by adding an attention module. The encoder and the improved decoder form the auxiliary branch of the model, which completes the segmentation task. The model proposed in this paper has two advantages. First, for small defects in images, the segmentation-based auxiliary task enables the model to focus on these small targets during the learning process and learn features with more representational power. Second, the introduction of an attention mechanism in the decoder can suppress the interference information, retain the effective location information, focus on the learning of defective regions, adjust the different feature mapping weights, and improve the performance of the model. The method has been validated on the self-built dataset and the dataset collected from X-ray inspection industrial sites and compared with typical image classification methods. The results show that the proposed method in this paper has higher defect recognition accuracy than other classical deep network models and can effectively identify multiple types of defect features while providing assurance and reference for the quality and safety performance of the parts to be inspected.