High-quality Datasets Research Articles

Real-time instance segmentation of recyclables from highly cluttered construction and demolition waste streams.

Efficient recycling strategies are crucial for mitigating the adverse environmental impacts of escalating construction and demolition waste (CDW). While automated identification via deep-learning is a promising direction, localizing CDW recyclables is uniquely challenging due to significant clutter and compositional complexity. Recognizing that accurate and fast localization is a strong prerequisite for swift robotic action, this study provides a comprehensive assessment of state-of-the-art (s.o.t.a) real-time instance segmentation of recyclables from complex CDW streams. Unlike previous studies that simplify the real-world conditions, this study curates and employs a high-quality CDW instance segmentation dataset tailored to capture often-ignored domain intricacies such as deformation, contamination, intricate class distinctions, scale variations, and high levels of clutter, ensuring practical applicability. The results show that even advanced networks struggle with complexity of real CDW streams due to high clutter, with segmentation accuracy not surpassing 50%. To address this, the study proposes integrating patch-based inferencing techniques, which help the model focus on cluttered regions more effectively, boosting overall performance by a notable 12.9%. Additionally, to enhance the zero-shot capabilities of s.o.t.a prompt-based real-time segmentation for identifying CDW recyclables, a simple yet effective domain-transfer framework is proposed, substantially increasing number of correct mask predictions by 68%. This study serves as a practical reference for applying deep-learning tools in automated environmental management tasks such as waste sorting and suggests the best architectural framework for practical use.

A multi-source heterogeneous medical data enhancement framework based on lakehouse.

Obtaining high-quality data sets from raw data is a key step before data exploration and analysis. Nowadays, in the medical domain, a large amount of data is in need of quality improvement before being used to analyze the health condition of patients. There have been many researches in data extraction, data cleaning and data imputation, respectively. However, there are seldom frameworks integrating with these three techniques, making the dataset suffer in accuracy, consistency and integrity. In this paper, a multi-source heterogeneous data enhancement framework based on a lakehouse MHDP is proposed, which includes three steps of data extraction, data cleaning and data imputation. In the data extraction step, a data fusion technique is offered to handle multi-modal and multi-source heterogeneous data. In the data cleaning step, we propose HoloCleanX, which provides a convenient interactive procedure. In the data imputation step, multiple imputation (MI) and the SOTA algorithm SAITS, are applied for different situations. We evaluate our framework via three tasks: clustering, classification and strategy prediction. The experimental results prove the effectiveness of our data enhancement framework.

High-fidelity prediction of forming quality for self-piercing riveted joints in aluminum alloy based on machine learning

The effect of rivet length and sheet thickness on the cross-sectional formation and tensile-shear performance of self-piercing riveted joints in AA5754 aluminum alloy was examined through experimental investigation. The influence degree of joining parameters on the forming quality was analyzed. It was revealed that rivet length and sheet thickness are pivotal factors influencing the tensile-shear strength of the joint, culminating in the identification of four optimal riveting process schemes:LCh1Ah2A、LAh1Ah2B、LAh1Ah2B and LBh1Ah2B. A simulation model for self-piercing riveting was established, employing the GISSMO failure model and the modified Mohr-Coulomb (MMC) failure criteria to predict the damage and fracture of the aluminum alloy. A plethora of high-quality datasets depicting the cross-sections of the joints were derived from simulation analysis. Subsequently, the structure and hyperparameter determination method of traditional neural network prediction models were elucidated. By amalgamating the Aquila Optimization (AO) algorithm with the African Vultures Optimization Algorithm (AVOA), a hybrid optimization algorithm model known as MIC_AOAVOA was developed. This model effectively harnesses the strengths of various algorithms to augment search efficiency and optimization capabilities. Strategies for population initialization and adaptive weight adjustments were incorporated to enhance the algorithm's convergence velocity and the quality of solutions. The cauchy opposition-based learning (COBL) and fitness-distance balance (FDB) strategy further refined the composite algorithm, bolstering its global search capabilities and population diversity. Comparative analyses were performed with single algorithm models and traditional BP neural network models, with an in-depth examination of the MIC_AOAVOA_BP model's prediction outcomes. Comprehensive evaluations utilizing error statistics and composite evaluation indicators demonstrated that the model consistently achieved mean absolute percentage error (MAPE) values below 10%, correlation coefficients (R²) exceeding 0.98, and stable mean squared error (MSE) values around 0.0002 across the prediction of three metrics. These results underscore the model's high precision and stability. Consequently, the proposed enhanced model offers a solution that is more stable, accurate, and robust for the prediction of forming quality in self-piercing riveted joints within engineering applications.

Suitability of large language models for extraction of high-quality chemical reaction dataset from patent literature.

With the advent of artificial intelligence (AI), it is now possible to design diverse and novel molecules from previously unexplored chemical space. However, a challenge for chemists is the synthesis of such molecules. Recently, there have been attempts to develop AI models for retrosynthesis prediction, which rely on the availability of a high-quality training dataset. In this work, we explore the suitability of large language models (LLMs) for extraction of high-quality chemical reaction data from patent documents. A comparative study on the same set of patents from an earlier study showed that the proposed automated approach can enhance the current datasets by addition of 26% new reactions. Several challenges were identified during reaction mining, and for some of them alternative solutions were proposed. A detailed analysis was also performed wherein several wrong entries were identified in the previously curated dataset. Reactions extracted using the proposed pipeline over a larger patent dataset can improve the accuracy and efficiency of synthesis prediction models in future.Scientific contributionIn this work we evaluated the suitability of large language models for mining a high-quality chemical reaction dataset from patent literature. We showed that the proposed approach can significantly improve the quantity of the reaction database by identifying more chemical reactions and improve the quality of the reaction database by correcting previous errors/false positives.

Research and Application of YOLOv11-Based Object Segmentation in Intelligent Recognition at Construction Sites

With the increasing complexity of construction site environments, robust object detection and segmentation technologies are essential for enhancing intelligent monitoring and ensuring safety. This study investigates the application of YOLOv11-Seg, an advanced target segmentation technology, for intelligent recognition on construction sites. The research focuses on improving the detection and segmentation of 13 object categories, including excavators, bulldozers, cranes, workers, and other equipment. The methodology involves preparing a high-quality dataset through cleaning, annotation, and augmentation, followed by training the YOLOv11-Seg model over 351 epochs. The loss function analysis indicates stable convergence, demonstrating the model’s effective learning capabilities. The evaluation results show an mAP@0.5 average of 0.808, F1 Score(B) of 0.8212, and F1 Score(M) of 0.8382, with 81.56% of test samples achieving confidence scores above 90%. The model performs effectively in static scenarios, such as equipment detection in Xiong’an New District, and dynamic scenarios, including real-time monitoring of workers and vehicles, maintaining stable performance even at 1080P resolution. Furthermore, it demonstrates robustness under challenging conditions, including nighttime, non-construction scenes, and incomplete images. The study concludes that YOLOv11-Seg exhibits strong generalization capability and practical utility, providing a reliable foundation for enhancing safety and intelligent monitoring at construction sites. Future work may integrate edge computing and UAV technologies to support the digital transformation of construction management.

Wi-Fi Sensing Techniques for Human Activity Recognition: Brief Survey, Potential Challenges, and Research Directions

Recent advancements in wireless communication technologies have made Wi-Fi signals indispensable in both personal and professional settings. The utilization of these signals for Human Activity Recognition (HAR) has emerged as a cutting-edge technology. By leveraging the fluctuations in Wi-Fi signals for HAR, this approach offers enhanced privacy compared to traditional visual surveillance methods. The essence of this technique lies in detecting subtle changes when Wi-Fi signals interact with the human body, which are then captured and interpreted by advanced algorithms. This paper initially provides an overview of the key methodologies in HAR and the evolution of non-contact sensing, introducing sensor-based recognition, computer vision, and Wi-Fi signal-based approaches, respectively. It then explores tools for Wi-Fi-based HAR signal collection and lists several high-quality datasets. Subsequently, the paper reviews various sensing tasks enabled by Wi-Fi signal recognition, highlighting the application of deep learning networks in Wi-Fi signal detection. The fourth section presents experimental results that assess the capabilities of different networks. The findings indicate significant variability in the generalization capacities of neural networks and notable differences in test accuracy for various motion analyses.

Influence of 3D Barriers on Walkability for the Elderly in a German City

Walking is a sustainable, safe, and active mode of transportation. The benefits that the elderly in particular gain from outdoor walking are manifold, be it free andindependent access to stores and services of all kinds or the opportunity to socialize, enjoy parks, et cetera. This article depicts one particular factor that affects outdoor walkability,namely the gradient of walking paths. Steep slopes can be a serious obstacle to walkability, primarily for older people. The evaluation of available geospatial data sources formedthe basis for a geospatial analysis of walkability in the larger city of Kaiserslautern, located in southwest Germany. The concept of Walk Score was used to quantify the resultsobtained. The results demonstrate that the Walk Score can be refined to better address the mobility needs of older adults. The methodology was implemented for the German cityof Kaiserslautern by integrating volunteered geographic information with high-quality official datasets.

Leveraging machine learning in limited sampling strategies for efficient estimation of the area under the curve in pharmacokinetic analysis: a review.

Limited sampling strategies are widely employed in clinical practice to minimize the number of blood samples required for the accurate area under the curve calculations, as obtaining these samples can be costly and challenging. Traditionally, the maximum a posteriori Bayesian estimation has been the standard method for the area under the curve estimation based on limited samples. However, machine learning is emerging as a promising alternative for this purpose. Here, we review studies that utilize machine learning approaches to develop limited sampling strategies and compare the strengths and weaknesses of these machine learning methods. We searched the literature for studies that used machine learning to estimate the area under the curve using a limited sampling strategy approach. We identified ten studies that developed machine learning models to estimate the area under the curve for six different drugs. Several of these models demonstrated good accuracy and precision in area under the curve estimation in reference to the traditional Bayesian approach, highlighting the potential of machine learning models in precision dosing. Despite these promising early results, the development of machine learning for limited sampling strategies is still in its early stages. Further research might be needed to validate machine learning models with larger, high-quality clinical datasets to ensure their reliability and applicability in clinical settings.

DeFFace: Deep Face Recognition Unlocked by Illumination Attributes

General face recognition is currently one of the key technologies in the field of computer vision, and it has achieved tremendous success with the support of deep-learning technology. General face recognition models currently exhibit extremely high accuracy on some high-quality face datasets. However, their performance decreases in challenging environments, such as low-light scenes. To enhance the performance of face recognition models in low-light scenarios, we propose a face recognition approach based on feature decoupling and fusion (DeFFace). Our main idea is to extract facial-related features from images that are not influenced by illumination. First, we introduce a feature decoupling network (D-Net) to decouple the image into facial-related features and illumination-related features. By incorporating the illumination triplet loss optimized with unpaired identity IDs, we regulate illumination-related features to minimize the impact of lighting conditions on the face recognition system. However, the decoupled features are relatively coarse. Therefore, we introduce a feature fusion network (F-Net) to further extract the residual facial-related features from the illumination-related features and fuse them with the initial facial-related features. Finally, we introduce a lighting-facial correlation loss to reduce the correlation between the two decoupled features in the specific space. We demonstrate the effectiveness of our method on four real-world low-light datasets and three simulated low-light datasets. We retrain multiple general face recognition methods using our proposed low-light training sets to further validate the advanced performance of our method. Compared to general face recognition methods, our approach achieves an average improvement of more than 2.11 percentage points on low-light face datasets. In comparison with image enhancement-based solutions, our method shows an average improvement of around 16 percentage points on low-light datasets, and it also delivers an average improvement of approximately 5.67 percentage points when compared to illumination normalization-based methods.

A Dataset of 3M Single-DOF Planar 4-, 6-, and 8-Bar Linkage Mechanisms With Open and Closed Coupler Curves for Machine Learning-Driven Path Synthesis

Abstract In recent years, there has been a strong interest in applying machine learning techniques to path synthesis of linkage mechanisms. However, progress has been stymied due to a scarcity of high-quality datasets. In this article, we present a comprehensive dataset comprising nearly three million samples of 4-, 6-, and 8-bar linkage mechanisms with open and closed coupler curves. Current machine learning approaches to path synthesis also lack standardized metrics for evaluating outcomes. To address this gap, we propose six key metrics to quantify results, providing a foundational framework for researchers to compare new models with existing ones. We also present a variational autoencoder-based model in conjunction with a k-nearest neighbor search approach to demonstrate the utility of our dataset. In the end, we provide example mechanisms that generate various curves along with a numerical evaluation of the proposed metrics.

Can Linguistic Knowledge Improve Multimodal Alignment in Vision-Language Pretraining?

The field of multimedia research has witnessed significant interest in leveraging multimodal pretrained neural network models to perceive and represent the physical world. Among these models, vision-language pretraining (VLP) has emerged as a captivating topic. Currently, the prevalent approach in VLP involves supervising the training process with paired image-text data. However, limited efforts have been dedicated to exploring the extraction of essential linguistic knowledge, such as semantics and syntax, during VLP and understanding its impact on multimodal alignment. In response, our study aims to shed light on the influence of comprehensive linguistic knowledge encompassing semantic expression and syntactic structure on multimodal alignment. To achieve this, we introduce SNARE , a large-scale multimodal alignment probing benchmark designed specifically for the detection of vital linguistic components, including lexical, semantic, and syntax knowledge. SNARE offers four distinct tasks: Semantic Structure, Negation Logic, Attribute Ownership, and Relationship Composition. Leveraging SNARE , we conduct holistic analyses of six advanced VLP models (BLIP, CLIP, Flava, X-VLM, BLIP2, and GPT-4), along with human performance, revealing key characteristics of the VLP model: (i) Insensitivity to complex syntax structures, relying primarily on content words for sentence comprehension. (ii) Limited comprehension of sentence combinations and negations. (iii) Challenges in determining actions or spatial relations within visual information, as well as difficulties in verifying the correctness of ternary relationships. Based on these findings, we propose the following strategies to enhance multimodal alignment in VLP: (1) Utilize a large generative language model as the language backbone in VLP to facilitate the understanding of complex sentences. (2) Establish high-quality datasets that emphasize content words and employ simple syntax, such as short-distance semantic composition, to improve multimodal alignment. (3) Incorporate more fine-grained visual knowledge, such as spatial relationships, into pretraining objectives. 1

Advancing clinical MRI exams with artificial intelligence: Japan's contributions and future prospects.

In this narrative review, we review the applications of artificial intelligence (AI) into clinical magnetic resonance imaging (MRI) exams, with a particular focus on Japan's contributions to this field. In the first part of the review, we introduce the various applications of AI in optimizing different aspects of the MRI process, including scan protocols, patient preparation, image acquisition, image reconstruction, and postprocessing techniques. Additionally, we examine AI's growing influence in clinical decision-making, particularly in areas such as segmentation, radiation therapy planning, and reporting assistance. By emphasizing studies conducted in Japan, we highlight the nation's contributions to the advancement of AI in MRI. In the latter part of the review, we highlight the characteristics that make Japan a unique environment for the development and implementation of AI in MRI examinations. Japan's healthcare landscape is distinguished by several key factors that collectively create a fertile ground for AI research and development. Notably, Japan boasts one of the highest densities of MRI scanners per capita globally, ensuring widespread access to the exam. Japan's national health insurance system plays a pivotal role by providing MRI scans to all citizens irrespective of socioeconomic status, which facilitates the collection of inclusive and unbiased imaging data across a diverse population. Japan's extensive health screening programs, coupled with collaborative research initiatives like the Japan Medical Imaging Database (J-MID), enable the aggregation and sharing of large, high-quality datasets. With its technological expertise and healthcare infrastructure, Japan is well-positioned to make meaningful contributions to the MRI-AI domain. The collaborative efforts of researchers, clinicians, and technology experts, including those in Japan, will continue to advance the future of AI in clinical MRI, potentially leading to improvements in patient care and healthcare efficiency.

Prediction of automotive response using supervised machine learning in antilock braking systems and comparison with different models for improved vehicle safety

Abstract Previous research on predicting automotive response in antilock braking systems faced several common drawbacks. Limited availability of high-quality datasets and computational constraints hindered the models' ability to generalize and perform effectively in real-time. Additionally, integrating these models into existing vehicle systems is challenging, often requiring extensive calibration and testing to ensure reliability and safety. This study focuses on enhancing automotive safety through the integration of machine learning with advanced braking systems. An antilock braking system model is developed using MATLAB, accounting for key parameters such as speed, mass, wheel radius, and moment of inertia. Various probability ranges are simulated to evaluate the system's real-time response in dynamic environments. A supervised machine learning approach is applied, with lasso regression used for model validation. Braking response was evaluated at three speeds: 25 km/h, 45 km/h, and 65 km/h. To assess model accuracy, error analysis was conducted using metrics such as mean absolute percentage error (MAPE), mean absolute error (MAE), root mean squared error (RMSE), and mean squared error (MSE). Results showed a significant improvement in prediction accuracy, with MAPE reduced to 3.2%, MAE to 0.15 seconds, RMSE to 0.25 seconds, and MSE to 0.06 seconds.A comparative analysis is done using the Python Jupyter Framework, further validating the findings. The machine learning model successfully forecasts braking response times, and validates its prediction against random datasets.

Reconstruction of 30 m Land Cover in the Qilian Mountains from 1980 to 1990 Based on Super-Resolution Generative Adversarial Networks

Long time series of annual land cover with fine spatio-temporal resolutions play a crucial role in studying environmental climate change, biophysical modeling, carbon cycling models, and land management. Despite a strong consistency exhibited by several publicly available medium to fine resolution global land cover datasets, significant discrepancies exist at the regional scale; moreover, only every 5/10 year land cover were available. Consequently, high-quality annual land cover datasets before 2000 are unavailable in China. In this study, we proposed a deep learning-based method by integrating multiple remote sensing data from different platforms with historical high spatial resolution land cover datasets (CNLUCC) to derive the 30 m annual land cover maps from 1980 to 1990 for Qilian Mountain. First, the super-resolution generative adversarial network models for upscaling the 5.5 km AVHRR NDVI to 250 m were established by employing the AVHRR and MODIS NDVI data with the same year as input, and the early time series AVHRR NDVI data were subsequently upscaled to 250 m through the above models. Second, the breaks for the additive seasonal and trend (BFAST) change detection algorithm was applied to the upscaled time series NDVI data to detect the change time of different land cover types. Third, the CNLUCC data in 1980 and 1990 were updated to annual land cover datasets from 1980 to 1990 and the annual mapping results provided insights into the dynamic processes of urbanization, deforestation, water bodies, and farmland from 1980 to 1990. Finally, comprehensive analysis and validation were carried out for evaluation and an overall accuracy of 77.26% for the land cover product in 1986 was achieved.

Streaming Large-Scale Microscopy Data to a Supercomputing Facility.

Data management is a critical component of modern experimental workflows. As data generation rates increase, transferring data from acquisition servers to processing servers via conventional file-based methods is becoming increasingly impractical. The 4D Camera at the National Center for Electron Microscopy generates data at a nominal rate of 480 Gbit s-1 (87,000 frames s-1), producing a 700 GB dataset in 15 s. To address the challenges associated with storing and processing such quantities of data, we developed a streaming workflow that utilizes a high-speed network to connect the 4D Camera's data acquisition system to supercomputing nodes at the National Energy Research Scientific Computing Center, bypassing intermediate file storage entirely. In this work, we demonstrate the effectiveness of our streaming pipeline in a production setting through an hour-long experiment that generated over 10 TB of raw data, yielding high-quality datasets suitable for advanced analyses. Additionally, we compare the efficacy of this streaming workflow against the conventional file-transfer workflow by conducting a postmortem analysis on historical data from experiments performed by real users. Our findings show that the streaming workflow significantly improves data turnaround time, enables real-time decision-making, and minimizes the potential for human error by eliminating manual user interactions.

Analysis of the impact of deep learning know-how and data in modelling neonatal EEG

The performance gains achieved by deep learning models nowadays are mainly attributed to the usage of ever larger datasets. In this study, we present and contrast the performance gains that can be achieved via accessing larger high-quality datasets versus the gains that can be achieved from harnessing the latest deep learning architectural and training advances. Modelling neonatal EEG is particularly affected by the lack of publicly available large datasets. It is shown that greater performance gains can be achieved from harnessing the latest deep learning advances than using a larger training dataset when adopting AUC as a metric, whereas using AUC90 or AUC-PR as metrics greater performance gains are achieved from using a larger dataset than harnessing the latest deep learning advances. In all scenarios the best performance is obtained by combining both deep learning advances and larger datasets. A novel developed architecture is presented that outperforms the current state-of-the-art model for the task of neonatal seizure detection. A novel method to fine-tune the presented model towards site-specific settings based on pseudo labelling is also outlined. The code and the weights of the model are made publicly available for benchmarking future model performances for neonatal seizure detection.

Computing hematopoiesis plasticity in response to genetic mutations and environmental stimulations.

Cell plasticity (CP), describing a dynamic cell state, plays a crucial role in maintaining homeostasis during organ morphogenesis, regeneration, and trauma-to-repair biological process. Single-cell-omics datasets provide an unprecedented resource to empower CP analysis. Hematopoiesis offers fertile opportunities to develop quantitative methods for understanding CP. In this study, we generated high-quality lineage-negative single-cell RNA-sequencing datasets under various conditions and introduced a working pipeline named scPlasticity to interrogate naïve and disturbed plasticity of hematopoietic stem and progenitor cells with mutational or environmental challenges. Using embedding methods UMAP or FA, a continuum of hematopoietic development is visually observed in wild type where the pipeline confirms a low proportion of hybrid cells ( P hc , with bias range: 0.4∼0.6) on a transition trajectory. Upon Tet2 mutation, a driver of leukemia, or treatment of DSS, an inducer of colitis, P hc is increased and plasticity of hematopoietic stem and progenitor cells was enhanced. We prioritized several transcription factors and signaling pathways, which are responsible for P hc alterations. In silico perturbation suggests knocking out EGR regulons or pathways of IL-1R1 and β-adrenoreceptor partially reverses P hc promoted by Tet2 mutation and inflammation.

Enhancing sustainable Chinese cabbage production: a comparative analysis of multispectral image instance segmentation techniques

Accurate instance segmentation of individual crops is crucial for field management and crop monitoring in smart agriculture. To address the limitations of traditional remote sensing methods in individual crop analysis, this study proposes a novel instance segmentation approach combining UAVs with the YOLOv8-Seg model. The YOLOv8-Seg model supports independent segmentation masks and detection at different scales, utilizing Path Aggregation Feature Pyramid Networks (PAFPN) for multi-scale feature integration and optimizing sample matching through the Task-Aligned Assigner. We collected multispectral data of Chinese cabbage using UAVs and constructed a high-quality dataset via semi-automatic annotation with the Segment Anything Model (SAM). Using mAP as the evaluation metric, we compared YOLO series algorithms with other mainstream instance segmentation methods and analyzed model performance under different spectral band combinations and spatial resolutions. The results show that YOLOv8-Seg achieved 86.3% mAP under the RGB band and maintained high segmentation accuracy at lower spatial resolutions (1.33 ~ 1.14 cm/pixel), successfully extracting key metrics such as cabbage count and average leaf area. These findings highlight the potential of integrating UAV technology with advanced segmentation models for individual crop monitoring, supporting precision agriculture applications.

Systematic review and meta-analysis of bulk RNAseq studies in human Alzheimer's disease brain tissue.

To systematically review and meta-analyze bulk RNA sequencing studies comparing Alzheimer's disease (AD) patients with controls in human brain tissue, assessing study quality and identifying key genes and pathways. We searched PubMed, Web of Science, and Scopus on September 23, 2023, for studies using bulk RNAseq on primary human brain tissue from AD patients and controls. Excluded were non-primary tissue, re-analyses without new data, limited RNA types and gene panels. Quality was assessed with a 10-category tool. Meta-analysis used high-quality datasets. From 3,266 records, 24 studies met criteria. Meta-analysis found 571 differentially expressed genes (DEGs) in temporal lobe and 189 in frontal lobe; overlapping pathways included "Tube morphogenesis" and "Neuroactive ligand-receptor interaction." Study heterogeneity and limited data tables constrained the review. Rigorous methods are vital in AD transcriptomic studies. Findings enhance understanding of transcriptomic changes, aiding biomarker and therapeutic development. PROSPERO (CRD42023466522).

Clean-Label Backdoor Watermarking for Dataset Copyright Protection via Trigger Optimization

High-quality datasets are essential for training high-performance models, while the process of collection, cleaning, and labeling is costly. As a result, datasets are considered valuable intellectual property. However, when security mechanisms are symmetry-breaking, creating exploitable vulnerabilities, unauthorized use or data leakage can infringe on the copyright of dataset owners. In this study, we design a method to mount clean-label dataset watermarking based on trigger optimization, aiming to protect the copyright of the dataset from infringement. We first perform iterative optimization of the trigger based on a surrogate model, with targets class samples guiding the updates. The process ensures that the optimized triggers contain robust feature representations of the watermark target class. A watermarked dataset is obtained by embedding optimized triggers into randomly selected samples from the watermark target class. If an adversary trains a model with the watermarked dataset, our watermark will manipulate the model’s output. By observing the output of the suspect model on samples with triggers, it can be determined whether the model was trained on the watermarked dataset. The experimental results demonstrate that the proposed method exhibits high imperceptibility and strong robustness against pruning and fine-tuning attacks. Compared to existing methods, the proposed method significantly improves effectiveness at very low watermarking rates.