Labeling Data Research Articles

Classification of Subdivision Land Use and Land Cover Using Deep Learning Models

Land cover provides crucial information related to biological geography, ecological climatology, and human activities. In the past, land cover mapping was performed based on visual interpretation, but it had limitations in terms of time and cost. Recently, it has become possible to create land cover maps with higher temporal resolution over wider areas using artificial intelligence-based models. The accuracy and reliability of AI model-based land cover maps increase with the amount of training data, but it is difficult to acquire large amounts of data due to the time required for label data annotation. In South Korea, the Environmental Geographic Information Service provides self-learning data consisting of aerial orthoimages and subdivision land cover classification level label data, making it possible to collect high-quality data. Therefore, this study examined the feasibility of self-learning data by building and evaluating a U-Net-based land cover classification model for waterfront areas using self-learning data. The trained model showed relatively low performance with an F-1 score of 0.61 for training data and 0.31 for test data. The model’s low performance is thought to be due to insufficient training caused by the large number of classification categories (34) and data imbalance between categories. Although the model performance using self-learning data was low, it is believed that model performance can be improved by grouping classification categories according to research purposes or resolving data imbalance through data augmentation techniques. Therefore, self-learning data is expected to be utilized in various studies using land cover.

GeoAI Dataset for Urban Water Body Detection Using TerraSAR-X Satellite Radar Imagery

This study presents the generation of a GeoAI dataset for urban water body detection using TerraSAR-X satellite synthetic aperture radar (SAR) imagery. The study area includes urban regions in Seoul and Gyeonggi Province, chosen for their complex structures and frequent flooding, which pose challenges for SAR analysis. The data preprocessing involved generating Sigma0 images, image co-registration, median filtering for speckle noise reduction, decibel conversion, and orthorectification using Copernicus DEM for precise geometric correction. Label data were created using the global river widths from Landsat dataset combined with the Otsu thresholding method and fine-tuned with Google Map imagery. Annotation guidelines were meticulously designed to account for SAR-specific phenomena such as layover, corner reflections, and side lobe effects, ensuring consistent and accurate labeling across different orbits and observation conditions. The resulting dataset supports deep learning models in learning geometric characteristics of SAR imagery, enhancing water body detection capabilities. This work provides a foundational resource for future applications in urban water management and climate-resilient disaster response.

A Novel FS-GAN-Based Anomaly Detection Approach for Smart Manufacturing

In this study, we present the few-shot generative adversarial network (FS-GAN) model, which integrates few-shot learning and a generative adversarial network with an unsupervised learning approach (AnoGAN) to address the challenges of anomaly detection in smart-factory manufacturing environments. Manufacturing processes often encounter malfunctions or defective parts that disrupt production and compromise product quality. However, collecting and labeling sufficient data to detect anomalies is time-intensive, and abnormal data are rare, leading to data imbalances. The FS-GAN model leverages few-shot learning to enable accurate predictions with minimal data and uses the generative capabilities of AnoGAN to mitigate the scarcity of abnormal data by generating synthetic normal data. Experimental results demonstrate that FS-GAN outperforms existing models in terms of accuracy and learning speed, even with limited datasets, effectively addressing the data imbalance problem inherent in manufacturing. The model reduces dependency on extensive data collection and labeling efforts, making it suitable for real-world applications. Through reliable and efficient anomaly detection, FS-GAN contributes to production reliability, product quality, and operational efficiency in smart factories. This study highlights the potential of FS-GAN to provide a cost-effective and high-performance solution to the challenges of anomaly detection in the manufacturing industry.

GeoAI Dataset for Training a Deep Learning-based GEMS Snow Detection Model

The Geostationary Environment Monitoring Spectrometer (GEMS) observes air quality across East Asia from an altitude of approximately 36,000 km, analyzing the spatiotemporal distribution of atmospheric pollutants that spread beyond localized regions. GEMS currently provides 21 core air quality-related products, most of which are derived from Level 1C data, which has undergone geometric and radiometric correction. For enhanced accuracy in air quality analysis, precise surface reflectance estimation is essential. However, high-reflectance elements, such as snow, interfere with the accurate estimation of radiance values, necessitating precise detection of such areas. Despite this, GEMS relies solely on the ultraviolet and partial visible bands, lacking the infrared bands crucial for snow detection, and it has no proprietary snow detection algorithm, instead utilizing near-real-time ice and snow extent data from the U.S. National Snow and Ice Data Center. Recently, deep learning techniques have shown potential in image processing, outperforming traditional algorithms, which could address these limitations. However, there is currently no deep learning training dataset available for snow detection specifically for GEMS. To address this issue, this study developed a GeoAI dataset for training a deep learning-based snow detection model for GEMS. In this research, we constructed input data using GEMS Level 1C data and generated label data based on GEMS, Advanced Meteorological Imager, and MODIS snow cover data. The snow detection dataset developed in this study is expected to address the snow detection limitations of GEMS, providing foundational data to enhance the reliability of future geostationary satellite-based air quality research.

Wheat disease recognition method based on the SC-ConvNeXt network model

When utilizing convolutional neural networks for wheat disease identification, the training phase typically requires a substantial amount of labeled data. However, labeling data is both complex and costly. Additionally, the model’s recognition performance is often disrupted by complex factors in natural environments. To address these issues, this paper proposes a wheat disease identification model, SC-ConvNeXt, which integrates the SimCLR pre-training framework and an improved CBAM attention mechanism. The model employs ConvNeXt-T as the feature extraction network. Initially, it uses the self-supervised SimCLR pre-training framework to learn inter-class similarities, reducing the reliance on labeled data during training. Subsequently, the CBAM attention module is integrated into ConvNeXt-T to enhance the model’s feature extraction and generalization capabilities in complex backgrounds, and each attention module’s loss function is improved with a LeakyReLU activation function to prevent neuron deactivation when inputs are negative. Furthermore, by introducing the Focal Loss function, the model addresses the imbalance in the quantity of easy and difficult-to-classify samples. The dataset used in this study comes from the ‘Smart Agriculture’ platform of Jilin Agricultural Science and Technology College, including images of three wheat diseases and healthy wheat. After expanding the dataset with various data augmentation methods, the effectiveness of adding SimCLR and the attention mechanism was sequentially verified. Comparative experiments were also conducted against four classic classification models. The experimental results show that the proposed SC-ConvNeXt model achieves an average classification accuracy of 88.05% on the test set, the highest among all comparative models. The model does not require additional labeled data during training, demonstrating its effectiveness in enhancing wheat disease recognition performance under natural environmental conditions without the need for extra labeled data training.

RECOGNIZING A TRANSFORMER OIL GAUGE IN AN IMAGE USING A PRE-TRAINED MOBILENETV2 FPN LITE MODEL

Oil gauges allow you to monitor the amount of oil in transformers, which is necessary to prevent its overheating and related negative consequences, such as deterioration of the insulating and cooling properties of the oil, increase in pressure inside the device and possible accidents. Most of the oil pointers do not have an electric output to indicate the measurement results, which complicates the process of reading the readings of the device. One of the options for solving this problem may be the use of intelligent methods of image analysis, which have recently been actively used in many scientific and practical fields. The purpose of the study was to develop and evaluate the effectiveness of an intelligent system for determining the presence of an arrow transformer oil gauge in an image using a pre-trained MobileNetV2 FPN Lite model. Materials and methods. In the course of the study, theoretical and empirical methods were used. The study considers the pointer oil gauges of the MS-1 and MS-2 models. To develop a machine learning model for image recognition, the Python language was chosen, the open-source TensorFlow Object Detection API platform, which is part of TensorFlow, was used, which implements the creation, training, and testing of neural network object detection models. Research results. The process of generating a training set using augmentation methods and the process of data labeling are described. The process of selecting hyperparameters for model training was considered and analyzed. The hyperparameter “number of neurons” was used with the same value as in the pre-trained model. ReLU was chosen as the activation function, and the momentum optimization method was used. The batch size was 4, and the number of epochs was assumed to be 2000. The Accuracy, Recall, Precision, and F1-score metrics were calculated, and the quality and efficiency of the resulting model were evaluated based on them. The model did a good job of avoiding false positives, which confirms the high value of the Precision = 1.0 metric. However, it incorrectly recognized a significant number of non-oil gauge objects, as reflected by the low Recall = 0.687 value. The reason for such indicators of the Recall and F1-score metrics is the imbalance of data for training. To eliminate this drawback in further research, it is necessary to prepare a larger number of unique images of this device, displaying it at different angles of inclination, in different weather conditions, in different lighting. Conclusions. A model of an intelligent system for determining the presence of a transformer oil gauge in an image using a pre-trained MobileNetV2 FPN Lite model has been built. The model was trained to determine the device and its boundaries on the image, but the accuracy of the model is insufficient for its use in real conditions. Increasing the training, validation and test sample size can solve this problem. In general, the proposed method of recognizing transformer oil indicator on the image has shown its performance.

Deep Bayesian active learning using in-memory computing hardware

Labeling data is a time-consuming, labor-intensive and costly procedure for many artificial intelligence tasks. Deep Bayesian active learning (DBAL) boosts labeling efficiency exponentially, substantially reducing costs. However, DBAL demands high-bandwidth data transfer and probabilistic computing, posing great challenges for conventional deterministic hardware. Here we propose a memristor stochastic gradient Langevin dynamics in situ learning method that uses the stochastic of memristor modulation to learn efficiency, enabling DBAL within the computation-in-memory (CIM) framework. To prove the feasibility and effectiveness of the proposed method, we implemented in-memory DBAL on a memristor-based stochastic CIM system and successfully demonstrated a robot’s skill learning task. The inherent stochastic characteristics of memristors allow a four-layer memristor Bayesian deep neural network to efficiently identify and learn from uncertain samples. Compared with cutting-edge conventional complementary metal-oxide-semiconductor-based hardware implementation, the stochastic CIM system achieves a remarkable 44% boost in speed and could conserve 153 times more energy.

Techniques for Data Augmentation and Their Impact on Long-Range Dependence and Applications

Data augmentation is a common approach to enhance datasets for training machine learning models. This study employs five distinct techniques to generate augmented datasets. Furthermore, eight measures are applied to assess datasets both before and after augmentation techniques. A critical requirement is that any augmentation should preserve the fundamental properties of the original dataset. The study reveals that certain augmentation methods can disrupt the long-range dependence on Internet traffic data (ITD) with distributed denial of service (DDoS) attacks (DDoS ITD). These DDoS ITDs originate from stochastic and bursty environments, affecting the probability mass function (PMF) and data labeling.

Learning Semantic-Aware Representation in Visual-Language Models for Multi-Label Recognition with Partial Labels

Multi-label recognition with partial labels (MLR-PL), in which only some labels are known while others are unknown for each image, is a practical task in computer vision, since collecting large-scale and complete multi-label datasets is difficult in real application scenarios. Recently, vision language models (e.g. CLIP) have demonstrated impressive transferability to downstream tasks in data limited or label limited settings. However, current CLIP-based methods suffer from semantic confusion in MLR task due to the lack of fine-grained information in the single global visual and textual representation for all categories. In this work, we address this problem by introducing a semantic decoupling module and a category-specific prompt optimization method in CLIP-based framework. Specifically, the semantic decoupling module following the visual encoder learns category-specific feature maps by utilizing the semantic-guided spatial attention mechanism. Moreover, the category-specific prompt optimization method is introduced to learn text representations aligned with category semantics. Therefore, the prediction of each category is independent, which alleviate the semantic confusion problem. Extensive experiments on Microsoft COCO 2014 and Pascal VOC 2007 datasets demonstrate that the proposed framework significantly outperforms current state-of-art methods with a simpler model structure. Additionally, visual analysis shows that our method effectively separates information from different categories and achieves better performance compared to CLIP-based baseline method.

Low-Resource Chinese Named Entity Recognition via CNN-based Multitask Learning

Named entity recognition (NER) is a fundamental subtask for information extraction that aims to locate and classify named entities in unstructured text into predefined categories. Recently, large-scale language models (LLMs) have achieved SOTA performance on a variety of natural language processing tasks. However, because NER is a sequence labeling task in nature while LLMs is a text-generation model, the performance of LLMs on NER is still significantly below supervised baselines, and NER remains a difficult task. Meanwhile, the word boundary and semantic information of Chinese words are usually quite vague, as words contained in Chinese texts are not separated by spaces. Thus, the NER task still requires supervised learning paradigm and heavily relies on large amounts of labeled data, such as entity type and boundary information. However, the cost of labeling data can be prohibitively large, and the purely supervised approaches usually suffer from poor generalization capability. In this article, we propose a multitask learning-based bidirectional iterated dilated convolution model, BCNN-CWS, for low-resource NER via leveraging word boundary information of Chinese word segmentation (CWS) task. Specifically, to efficiently recognize named entities, an iterated dilated convolutional model with a limited number of layers is implemented. In addition, a bidirectional causal convolution mechanism is presented for contextual information extraction. Results of extensive experiments on public Chinese datasets demonstrate that BCNN-CWS achieves superior performance over state-of-the-art models, and it yields up to about 50% speed improvement over existing methods. It is worth noting that BCNN-CWS can be further improved by combining with a pretrained model. Received: 25 Spetember 2024 | Revised: 4 November 2024 | Accepted: 28 November 2024 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in GitLab at https://github.com/jiangfeng13/BCNN-CWS Author Contribution Statement Tao Wu: Conceptualization, Methodology, Writing – original draft, Writing – review & editing, Visualization, Supervision. Xinwen Cao: Resources, Data curation. Feng Jiang: Software, Validation, Formal analysis, Investigation, Writing – original draft. Canyixing Cui: Data curation, Writing -review & editing. Xuehao Li: Resources. Xingping Xian: Supervision, Project administration, Funding acquisition.

Improved self-training-based distant label denoising method for cybersecurity entity extractions.

The task of named entity recognition (NER) plays a crucial role in extracting cybersecurity-related information. Existing approaches for cybersecurity entity extraction predominantly rely on manual labelling data, resulting in labour-intensive processes due to the lack of a cybersecurity-specific corpus. In this paper, we propose an improved self-training-based distant label denoising method for cybersecurity entity extraction. Firstly, we create two domain dictionaries of cybersecurity. Then, an algorithm that combines reverse maximum matching and part-of-speech tagging restrictions is proposed, for generating distant labels for the cybersecurity domain corpus. Lastly, we propose a high-confidence text selection method and an improved self-training algorithm that incorporates a teacher-student model and weight update constraints, for exploring the true labels of low-confidence text using a model trained on high-confidence text, thereby reducing the noise in the distant annotation data. Experimental results demonstrate that the cybersecurity distantly-labelled data we obtained is of high quality. Additionally, the proposed constrained self-training algorithm effectively improves the F1 score of several state-of-the-art NER models on this dataset, yielding a 3.5% improvement for the Vendor class and a 3.35% improvement for the Product class.

Multi-Modal Prototypes for Few-Shot Object Detection in Remote Sensing Images

Few-shot object detection has attracted extensive attention due to the abomination of time-consuming or even impractical large-scale data labeling. Current studies attempted to employ prototype-matching approaches for object detection, constructing class prototypes from textual or visual features. However, single visual prototypes exhibit limited generalization in few-shot scenarios, while single textual prototypes lack the spatial details of remote sensing targets. Therefore, to achieve the best of both worlds, we propose a prototype aggregating module to integrate textual and visual prototypes, leveraging both semantics from textual prototypes and spatial details from visual prototypes. In addition, the transferability of multi-modal few-shot detectors from natural scenarios to remote sensing scenarios remains unexplored, and previous training strategies for FSOD do not adequately consider the characteristics of text encoders. To address the issue, we have conducted extensive ablation studies on different feature extractors of the detector and propose an efficient two-stage training strategy, which takes the characteristics of the text feature extractor into account. Experiments on two common few-shot detection benchmarks demonstrate the effectiveness of our proposed method. In four widely used data splits of DIOR, our method significantly outperforms previous state-of-the-art methods by at most 8.7%.

CACs Recognition of FISH Images Based on Adaptive Mean Teacher Semi-supervised Learning with Domain-Knowledge Pseudo Label.

Circulating genetically abnormal cells (CACs) serve as crucial biomarkers for lung cancer diagnosis. Detecting CACs holds great value for early diagnosis and screening of lung cancer. To aid the identification of CACs, we have incorporated deep learning algorithms into our CACs detection system, specifically developing algorithms for cell segmentation and signal point detection. However, it is noteworthy that deep learning algorithms require extensive data labeling. Consequently, this study introduces a semi-supervised learning algorithm for CACs detection. For the cell segmentation task, a combination of self-training and Mean Teacher method was adopted in the semi-supervised training cell segmentation task. Furthermore, an Adaptive Mean Teacher approach was developed based on the Mean Teacher to enhance the effectiveness of semi-supervised cell segmentation. Regarding the signal point detection task, an end-to-end semi-supervised signal point detection algorithm was developed using the Adaptive Mean Teacher as the paradigm, and a Domain-Knowledge Pseudo Label was developed to improve the quality of pseudo-labeling and further enhance signal point detection. By incorporating semi-supervised training in both sub-tasks, the reliance on labeled data is reduced, thereby improving the performance of CACs detection. Our proposed semi-supervised method has achieved good results in cell segmentation tasks, signal point detection tasks, and the final CACs detection task. In the final CACs detection task, with 2%, 5%, and 10% of labeled data, our proposed semi-supervised method achieved 27.225%, 23.818%, and 4.513%, respectively. Experimental results demonstrated that the proposed method is effective.

Real-Time Detection of IoT Anomalies and Intrusion Data in Smart Cities Using Multi-Agent System.

Analyzing IoT data is an important challenge in the smart cities domain due to the complexity of network traffic generated by a large number of interconnected devices: smart cameras, light bulbs, motion sensors, voice assistants, and so on. To overcome this issue, a multi-agent system is proposed to deal with all machine learning steps, from preprocessing and labeling data to discovering the most suitable model for the analyzed dataset. This paper shows that dividing the work into different tasks, managed by specialized agents, and evaluating the discovered models by an Expert System Agent leads to better results in the learning process.

Building robust malware detection through conditional Generative Adversarial Network-based data augmentation

Malware detection is essential in cybersecurity, yet its accuracy is often compromised by class imbalance and limited labeled data. This study leverages conditional Generative Adversarial Networks (cGANs) to generate synthetic malware samples, addressing these challenges by augmenting the minority class. The cGAN model generates realistic malware samples conditioned on class labels, balancing the dataset without altering the benign class. Applied to the CICMalDroid2020 dataset, the augmented data is used to train a LightGBM model, leading to improved detection accuracy, particularly for underrepresented malware classes. The results demonstrate the efficacy of cGANs as a robust data augmentation tool, enhancing the performance and reliability of machine learning-based malware detection systems.

Self Data Augmentation for Open Domain Question Answering

Information Retrieval (IR) constitutes a vital facet of Open-Domain Question Answering (ODQA) systems, focusing on the exploration of pertinent information within extensive collections of passages, such as Wikipedia, to facilitate subsequent reader processing. Historically, information retrieval relied on textual overlaps for relevant context retrieval, employing methods like BM25 and TF-IDF, which, however, lacked natural language understanding. The advent of deep learning ushered in a new era, leading to the introduction of Dense Passage Retrievers (DPR), shows superiority over traditional sparse retrievers. These dense retrievers leverage Pre-trained Language Models (PLMs) to initialize context encoders, enabling the extraction of natural language representations. They utilize the distance between latent vectors of contexts as a metric for assessing similarity. However, DPR methods are heavily reliant on large volumes of meticulously labeled data, such as Natural Questions. The process of data labeling is both costly and time-intensive. In this paper, we propose a novel data augmentation methodology SDA (Self Data Augmentation) that employs DPR models to automatically annotate unanswered questions. Specifically, we initiate the process by retrieving relevant pseudo passages for these unlabeled questions. We subsequently introduce three distinct passage selection methods to annotate these pseudo passages. Ultimately, we amalgamate the pseudo-labeled passages with the unanswered questions to create augmented data. Our experimental evaluations conducted on two extensive datasets (Natural Questions and TriviaQA), alongside a reletively small dataset (WebQuestions), utilizing three diverse base models, illustrate the significant enhancement achieved through the incorporation of freshly augmented data. Moreover, our proposed data augmentation method exhibits remarkable flexibility, which is readily adaptable to various dense retrievers. Additionally, we have conducted a comprehensive human study on the augmented data, which further supports our conclusions.

UltimateSynth: MRI Physics for Pan-Contrast AI.

Magnetic resonance imaging (MRI) is commonly used in healthcare for its ability to generate diverse tissue contrasts without ionizing radiation. However, this flexibility complicates downstream analysis, as computational tools are often tailored to specific MRI types and lack generalizability across the full spectrum of scans used in healthcare. Here, we introduce a versatile framework for the development and validation of pan-contrast AI models that can exhaustively cater to the full spectrum of scans achievable with MRI, enabling model deployment across scanner models, scan types, and age groups. Core to our framework is UltimateSynth, a technology that combines tissue physiology and MR physics in synthesizing realistic images across a comprehensive range of contrasts to bolster the AI development life cycle through efficient data labeling, generalizable model training, and thorough performance benchmarking. UltimateSynth is a platform for pan-contrast generalization of contrast-specific tools. We showcase the effectiveness of UltimateSynth by training an off-the-shelf U-Net to generalize anatomical segmentation across over 150,000 unique MRI contrasts, achieving robust tissue volumetric quantification with exceptionally low variability below 2%.

Automated Characterization of Abdominal MRI Exams Using Deep Learning.

Advances in magnetic resonance imaging (MRI) have revolutionized disease detection and treatment planning. However, as the volume and complexity of MRI data grow with increasing heterogeneity between institutions in imaging protocol, scanner technology, and data labeling, there is a need for a standardized methodology to efficiently identify, characterize, and label MRI sequences. Such a methodology is crucial for advancing research efforts that incorporate MRI data from diverse populations to develop robust machine learning models. This research utilizes convolutional neural networks (CNNs) to automatically classify sequence, orientation, and contrast, specifically tailored for abdominal MRI. Three distinct CNN models with similar backbone architectures were trained to classify single image slices into one of 12 sequences, 4 orientations, and 2 contrast classes. Results derived from this method demonstrate high levels of performance for the three specialized CNN models, with model accuracies for sequence, orientation, and contrast of 96.9%, 97.4%, and 97.3%, respectively.

Trace Copilot: Automatically Locating Cryptographic Operations in Side-Channel Traces by Firmware Binary Instrumenting

A common assumption in side-channel analysis is that the attacker knows the cryptographic algorithm implementation of the victim. However, many labsetting studies implicitly extend this assumption to the knowledge of the source code, by inserting triggers to measure, locate or align the Cryptographic Operations (CO) in the trace. For real-world attacks, the source code is typically unavailable, which poses a challenge for locating the COs thus reducing the effectiveness of many methods. In contrast, obtaining the (partial) binary firmware is more prevalent in practical attacks on embedded devices. While binary code theoretically encapsulates necessary information for side-channel attacks on software-implemented cryptographic algorithms, there is no systematic study on leveraging this information to facilitate side-channel analysis. This paper introduces a novel and general framework that utilizes binary information for the automated locating of COs on side-channel traces. We first present a mechanism that maps the execution flow of binary instructions onto the corresponding side-channel trace through a tailored static binary instrumentation process, thereby transforming the challenge of locating COs into one of tracing cryptographic code execution within the binary. For the latter, we propose a method to retrieve binary instruction addresses that are equivalent to the segmenting boundaries of the COs within side-channel traces. By identifying the mapping points of these instructions on the trace, we can obtain accurate segmentation labeling for the sidechannel data. Further, by employing the well-labeled side-channel segments obtained on a profiling device, we can readily identify the locations of COs within traces collected from un-controllable target devices. We evaluate our approach on various devices and cryptographic software, including a real-world secure boot program. The results demonstrate the effectiveness of our method, which can automatically locate typical COs, such as AES or ECDSA, in raw traces using only the binary firmware and a profiling device. Comparison experiments indicate that our method outperforms existing techniques in handling noisy or jittery traces and scales better to complex COs. Performance evaluation confirms that the runtime and storage overheads of the proposed approach are practical for real-world deployment.

PyPCG: a Python toolbox specialized for phonocardiography analysis.

Phonocardiography has recently gained popularity in low-cost and remote monitoring, including passive fetal heart monitoring. The development of methods which analyse phonocardiographic data tries to capitalize on this opportunity, and in recent years a multitude of such algorithms and models have been published. In these approaches there is little to no standardization and multiple parts of these models have to be reimplemented on a case-by-case basis. Datasets containing heart sound recordings also lack standardization in both data storage and labeling, especially in fetal phonocardiography. We are presenting a toolbox that can serve as a basis for a future standard framework for heart sound analysis. This toolbox contains some of the most widely used processing steps and with these, complex analysis pipelines can be created. These functions can be tested individually. Due to the interdependence of the steps, we validated the current segmentation stage using two phonocardiogram datasets, a fetal dataset comprising 50 one-minute abdominal PCG recordings, which include 6758 S1 and 6729 S2 labels and a filtered version of the dataset used in the 2022 PhysioNet Challenge, containing 413 records with 9795 S1 and 9761 S2 labels. Our results were compared to other common and publicly available segmentation methods, such as peak detection with the Neurokit2 library, and the Hidden Semi-Markov Model by Springer et al. Our best model achieved a 96.1% F1 score and 11.7 ms mean absolute error for fetal S1 detection, and 81.3% F1 score and 50.5 ms mean absolute error for PhysioNet S1 detection. Our detection method outperformed all other tested methods on the fetal dataset and
achieved results comparable to the state of the art on the PhysioNet dataset. Accurate segmentation of signals is critical for the calculation of accurate statistical measures and the creation of classification models. Our toolbox contains functions for both feature extraction and calculation of statistics which are compatible with the previous steps. All of our methods can be fine tuned for specific datasets. pyPCG is available on https://pypcg-toolbox.readthedocs.io/en/latest/.