Pub Date : 2025-12-11DOI: 10.1109/JSEN.2025.3640827
Kenesbaeva Periyzat Ismaylovna;Azimbek Khudoyberdiev;Hee-Cheol Kim
Traditional mental workload (MW) classification methods often rely on handcrafted features and achieve modest accuracy (70%–85%) while focusing on single modalities or static fusion, thus missing complementary information across sensors. Recent multimodal fusion approaches, such as attention-based weighting, averaging, or majority voting, often fail to accurately assess the relative informativeness of each modality, especially when one sensor becomes unreliable. We introduce CogniMoE, an end-to-end multimodal framework that learns from raw physiological signals with three innovations: 1) a high-efficiency on-the-fly scalogram generation pipeline using FP16 arithmetic that overcomes traditional storage bottlenecks reducing disk space usage by 98% while enabling seamless GPU processing; 2) parallel per-modality CNN–LSTM branches with attention and dynamic dropout that robustly extract modality-specific spatial–temporal features, outperforming single-stream encoders; and 3) an interpretable mixture of experts (MoE) gating mechanism that replaces static fusion with instance-level adaptive weighting, ensuring robustness by dynamically suppressing unreliable modalities in real time. Evaluations on the MAUS, CLAS, and WESAD datasets demonstrate that CogniMoE consistently outperforms both traditional methods (with average accuracies of 70%–85%) and recent state-ofthe- art (SOTA) approaches (up to 92% accuracy), achieving accuracies of 94%, 92%, and 98%, respectively. In addition, the MoE gating mechanism improves classification accuracy by approximately 5% on average over nonadaptive fusion strategies while dynamically adjusting modality importance based on individual participant characteristics.
{"title":"CogniMoE: End-to-End Multimodal Mental Workload Classification via On-the-Fly Scalogram Generation and MoE Gating","authors":"Kenesbaeva Periyzat Ismaylovna;Azimbek Khudoyberdiev;Hee-Cheol Kim","doi":"10.1109/JSEN.2025.3640827","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3640827","url":null,"abstract":"Traditional mental workload (MW) classification methods often rely on handcrafted features and achieve modest accuracy (70%–85%) while focusing on single modalities or static fusion, thus missing complementary information across sensors. Recent multimodal fusion approaches, such as attention-based weighting, averaging, or majority voting, often fail to accurately assess the relative informativeness of each modality, especially when one sensor becomes unreliable. We introduce CogniMoE, an end-to-end multimodal framework that learns from raw physiological signals with three innovations: 1) a high-efficiency on-the-fly scalogram generation pipeline using FP16 arithmetic that overcomes traditional storage bottlenecks reducing disk space usage by 98% while enabling seamless GPU processing; 2) parallel per-modality CNN–LSTM branches with attention and dynamic dropout that robustly extract modality-specific spatial–temporal features, outperforming single-stream encoders; and 3) an interpretable mixture of experts (MoE) gating mechanism that replaces static fusion with instance-level adaptive weighting, ensuring robustness by dynamically suppressing unreliable modalities in real time. Evaluations on the MAUS, CLAS, and WESAD datasets demonstrate that CogniMoE consistently outperforms both traditional methods (with average accuracies of 70%–85%) and recent state-ofthe- art (SOTA) approaches (up to 92% accuracy), achieving accuracies of 94%, 92%, and 98%, respectively. In addition, the MoE gating mechanism improves classification accuracy by approximately 5% on average over nonadaptive fusion strategies while dynamically adjusting modality importance based on individual participant characteristics.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 3","pages":"5213-5228"},"PeriodicalIF":4.3,"publicationDate":"2025-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11298422","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082139","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-18DOI: 10.1109/JSEN.2025.3631864
Naga Srinivasarao Chilamkurthy;Shaik Abdul Hakeem;Sreenivasulu Tupakula;Sunil Chinnadurai;Om Jee Pandey;Anirban Ghosh
The rapid expansion of Internet of Things (IoT) applications has driven advancements in networking technologies such as low-power wide-area networks (LPWANs) to extend coverage and enhance the lifespan of IoT devices (IoDs). However, real-world IoT networks are typically heterogeneous, comprising static and dynamic IoDs leading to variations in network topology. These fluctuations cause challenges such as increased data latency and energy imbalances, which hinder efficient information flow. To overcome these issues, this article presents a novel approach that integrates small-world characteristics (SWCs), inspired by social network theory, into heterogeneous LPWANs using reinforcement learning (RL). Specifically, the Q-learning technique is used to introduce new long-range links into the network, enhancing connectivity and optimizing performance. Different conventional networks with varying numbers of mobile nodes are studied in this work followed by their subsequent transformation to small-world versions. The performance of the networks is optimized in terms of energy efficiency and latency in data routing. It is observed that irrespective of the network (conventional or small-world), the performance is better if the number of static nodes is greater. Furthermore, independent of the degree of dynamicity, the SW-LPWAN is more energy-efficient and has lower transmission delay than the corresponding conventional network. Numerically, SWLPWANs achieve up to 14.6% faster data transmission speeds, supporting 19.7% more active IoDs, and maintaining 15.5% higher residual energy compared with conventional networks.
{"title":"Improving the Performance of Heterogeneous LPWANs: An Integrated Small-World and Machine Learning Approach","authors":"Naga Srinivasarao Chilamkurthy;Shaik Abdul Hakeem;Sreenivasulu Tupakula;Sunil Chinnadurai;Om Jee Pandey;Anirban Ghosh","doi":"10.1109/JSEN.2025.3631864","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3631864","url":null,"abstract":"The rapid expansion of Internet of Things (IoT) applications has driven advancements in networking technologies such as low-power wide-area networks (LPWANs) to extend coverage and enhance the lifespan of IoT devices (IoDs). However, real-world IoT networks are typically heterogeneous, comprising static and dynamic IoDs leading to variations in network topology. These fluctuations cause challenges such as increased data latency and energy imbalances, which hinder efficient information flow. To overcome these issues, this article presents a novel approach that integrates small-world characteristics (SWCs), inspired by social network theory, into heterogeneous LPWANs using reinforcement learning (RL). Specifically, the Q-learning technique is used to introduce new long-range links into the network, enhancing connectivity and optimizing performance. Different conventional networks with varying numbers of mobile nodes are studied in this work followed by their subsequent transformation to small-world versions. The performance of the networks is optimized in terms of energy efficiency and latency in data routing. It is observed that irrespective of the network (conventional or small-world), the performance is better if the number of static nodes is greater. Furthermore, independent of the degree of dynamicity, the SW-LPWAN is more energy-efficient and has lower transmission delay than the corresponding conventional network. Numerically, SWLPWANs achieve up to 14.6% faster data transmission speeds, supporting 19.7% more active IoDs, and maintaining 15.5% higher residual energy compared with conventional networks.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 1","pages":"1410-1419"},"PeriodicalIF":4.3,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145852503","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-17DOI: 10.1109/JSEN.2025.3631078
Lismer Andres Caceres Najarro;Iickho Song;Muhammad Salman;Kiseon Kim
In cooperative localization problems of wireless sensor networks (WSNs) these days, the estimation of target node (TN) positions is a challenging issue as the cost function becomes increasingly nonlinear and nonconvex due to the heightened interaction among sensor nodes. Most of the existing cooperative localization algorithms provide the acceptable localization accuracy, but with dramatically increased computational complexity. To reduce the computational complexity while maintaining competitive localization accuracy at the same time, we propose a localization algorithm based on the differential evolution (DE) with multiple populations, opposite-based learning, redirection, and anchoring. In the proposed scheme, the cost function is split into several simpler ones, each of which accounts only for one TN and is solved with a dedicated population. An enhanced version, which incorporates the population midpoint scheme, is also considered for further improvement in the localization accuracy. Simulation results demonstrate that the proposed algorithms provide comparable localization accuracy with much lower computational complexity against the state-of-the-art algorithms. In particular, the proposed algorithms reduce the execution time by up to 85% compared with other methods based on the semidefinite and second-order cone programming (SOCP), while delivering significantly higher localization accuracy than the faster yet less accurate least squares-based method.
{"title":"Multipopulation Differential Evolution for RSS-Based Cooperative Localization in Wireless Sensor Networks With Limited Communication Range","authors":"Lismer Andres Caceres Najarro;Iickho Song;Muhammad Salman;Kiseon Kim","doi":"10.1109/JSEN.2025.3631078","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3631078","url":null,"abstract":"In cooperative localization problems of wireless sensor networks (WSNs) these days, the estimation of target node (TN) positions is a challenging issue as the cost function becomes increasingly nonlinear and nonconvex due to the heightened interaction among sensor nodes. Most of the existing cooperative localization algorithms provide the acceptable localization accuracy, but with dramatically increased computational complexity. To reduce the computational complexity while maintaining competitive localization accuracy at the same time, we propose a localization algorithm based on the differential evolution (DE) with multiple populations, opposite-based learning, redirection, and anchoring. In the proposed scheme, the cost function is split into several simpler ones, each of which accounts only for one TN and is solved with a dedicated population. An enhanced version, which incorporates the population midpoint scheme, is also considered for further improvement in the localization accuracy. Simulation results demonstrate that the proposed algorithms provide comparable localization accuracy with much lower computational complexity against the state-of-the-art algorithms. In particular, the proposed algorithms reduce the execution time by up to 85% compared with other methods based on the semidefinite and second-order cone programming (SOCP), while delivering significantly higher localization accuracy than the faster yet less accurate least squares-based method.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 5","pages":"7770-7779"},"PeriodicalIF":4.3,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147299509","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-17DOI: 10.1109/JSEN.2025.3631094
Sihan Chen;Hui Chen;Shuqi Liu;Yige Zhao;Wanquan Liu
Outdoor depth acquisition with technologies like Light Detection and Ranging (LiDAR) is a challenging task due to factors such as high complexity and sensitivity to light variation, which result in sparse point cloud density. This research is an attempt to address these issues and suggests the use of Red, Green, Blue (RGB) image guidance for depth completion of sparse laser point clouds. The presented research work involves three stages. First, to overcome incomplete or inaccurate completion caused by unclear information corresponding RGB image and depth image, a guided convolutional module and a two-stage attention mechanism based on a feature fusion strategy are proposed. The strategy uses lightweight network models to improve the completion accuracy. Second, a completion method based on the fine-grained convolutional space propagation network is proposed to preserve the original depth value and refine the depth map. This scheme addresses the issue of losing the original depth value due to the noise while fusing two different information input modes of RGB image and depth map. Finally, in order to test the depth completion performance of TFDCNet, evaluation is performed by using the KITTI dataset. Experimental results reveal that TFDCNet shows improved completion accuracy by 8.36% in the selected scenarios compared with the state-of-the-art.
{"title":"TFDCNet: Two-Stage Multimodal Fusion and Fine-Grained Convolutional Space Propagation Network for Depth Completion of Outdoor Scenes","authors":"Sihan Chen;Hui Chen;Shuqi Liu;Yige Zhao;Wanquan Liu","doi":"10.1109/JSEN.2025.3631094","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3631094","url":null,"abstract":"Outdoor depth acquisition with technologies like Light Detection and Ranging (LiDAR) is a challenging task due to factors such as high complexity and sensitivity to light variation, which result in sparse point cloud density. This research is an attempt to address these issues and suggests the use of Red, Green, Blue (RGB) image guidance for depth completion of sparse laser point clouds. The presented research work involves three stages. First, to overcome incomplete or inaccurate completion caused by unclear information corresponding RGB image and depth image, a guided convolutional module and a two-stage attention mechanism based on a feature fusion strategy are proposed. The strategy uses lightweight network models to improve the completion accuracy. Second, a completion method based on the fine-grained convolutional space propagation network is proposed to preserve the original depth value and refine the depth map. This scheme addresses the issue of losing the original depth value due to the noise while fusing two different information input modes of RGB image and depth map. Finally, in order to test the depth completion performance of TFDCNet, evaluation is performed by using the KITTI dataset. Experimental results reveal that TFDCNet shows improved completion accuracy by 8.36% in the selected scenarios compared with the state-of-the-art.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 1","pages":"1395-1409"},"PeriodicalIF":4.3,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145852467","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Flexible sensors play a crucial role in intelligent wearable devices and flexible electronics. However, most pressure sensors focus primarily on enhancing sensitivity and detection range, overlooking the critical application requirements for low-detection limit (LOD) flexible pressure sensors in micro-pressure detection. Introducing internal microstructures has emerged as a promising strategy to enhance the micro-pressure response of flexible piezoresistive sensors. In this study, we employed an electrospun thermoplastic polyurethane fiber membrane featuring an internally hollow, vine-like microstructure as the flexible substrate. Carbon nanotubes were subsequently deposited onto the substrate via ultrasonication, resulting in a miniaturized flexible piezoresistive sensor with high sensitivity to micro-pressure. The nanofiber-based pressure sensor exhibited a high sensitivity of 23.264 Pa⁻¹ within the 0-5 Pa pressure range, alongside a thin profile (210 μm), rapid response/recovery times (60/60 ms), excellent LOD (0.5 Pa), and robust cycling stability over 3,000 continuous compression cycles. This sensor effectively monitors minute pressure variations and pulse signals, demonstrating significant potential for applications in intelligent wearable devices and flexible electronics.
{"title":"Hollow Vine-Like Thermoplastic Polyurethane-Based Sensor for Pulse Signal Detection in Motion State Classification Systems","authors":"Lingjie Kong, Jian Zhang, Xiaojing Yang, Yanjun Guo, Chi Zhang, Bokai Zhang, Ying Wang, Renhan Li, Yafei Qin","doi":"10.1109/jsen.2025.3630128","DOIUrl":"https://doi.org/10.1109/jsen.2025.3630128","url":null,"abstract":"Flexible sensors play a crucial role in intelligent wearable devices and flexible electronics. However, most pressure sensors focus primarily on enhancing sensitivity and detection range, overlooking the critical application requirements for low-detection limit (LOD) flexible pressure sensors in micro-pressure detection. Introducing internal microstructures has emerged as a promising strategy to enhance the micro-pressure response of flexible piezoresistive sensors. In this study, we employed an electrospun thermoplastic polyurethane fiber membrane featuring an internally hollow, vine-like microstructure as the flexible substrate. Carbon nanotubes were subsequently deposited onto the substrate via ultrasonication, resulting in a miniaturized flexible piezoresistive sensor with high sensitivity to micro-pressure. The nanofiber-based pressure sensor exhibited a high sensitivity of 23.264 Pa⁻¹ within the 0-5 Pa pressure range, alongside a thin profile (210 μm), rapid response/recovery times (60/60 ms), excellent LOD (0.5 Pa), and robust cycling stability over 3,000 continuous compression cycles. This sensor effectively monitors minute pressure variations and pulse signals, demonstrating significant potential for applications in intelligent wearable devices and flexible electronics.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 1","pages":"142-149"},"PeriodicalIF":0.0,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147330753","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-13DOI: 10.1109/JSEN.2025.3630222
Yi Ding;Zhili Zhang;Huayin Zhao;Jinyuan He;Liudi Wang;Jin Kang;Peng Xue;Wentao Cui;Enqing Dong
This article focuses on energy efficiency (EE) optimization for large-scale uniform line-distributed wireless geophone networks (WGNs). To address the issues of unbalanced subnetwork load and excessive energy consumption in traditional WGNs, we propose an EE optimization scheme—Markov-chain-based clustering and farthest vector forwarding (MCBC–FVF). For intracluster energy optimization, a state transition probability model based on Markov chain (MC) is constructed for cluster head (CH) election. An energy-aware objective function with a spatial bias term is designed to reduce and balance energy consumption. For intercluster energy optimization, a farthest vector forwarding (FVF) mechanism is introduced to mitigate communication failure and low packet delivery ratio (PDR) caused by excessive distances between CHs. It also helps reduce redundant traffic and suppresses path inflation. Compared with MH-LEACH, LEACH-C, and MMRP, simulation results based on IEEE 802.15.4 demonstrate that the proposed MCBC–FVF scheme improves the first node death (FND) time by 38.68%, 22.62%, and 17.20%, respectively, while reducing intercluster average energy consumption by 32.73%, 18.41%, and 43.38%, respectively. These results indicate that MCBC–FVF not only significantly prolongs network lifetime but also provides a novel integration of probabilistic modeling and topologyaware forwarding, offering a practical and effective solution for energy-constrained WGNs.
{"title":"An Energy Efficiency Optimization Scheme for Uniform Line-Distributed Wireless Geophone Networks","authors":"Yi Ding;Zhili Zhang;Huayin Zhao;Jinyuan He;Liudi Wang;Jin Kang;Peng Xue;Wentao Cui;Enqing Dong","doi":"10.1109/JSEN.2025.3630222","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3630222","url":null,"abstract":"This article focuses on energy efficiency (EE) optimization for large-scale uniform line-distributed wireless geophone networks (WGNs). To address the issues of unbalanced subnetwork load and excessive energy consumption in traditional WGNs, we propose an EE optimization scheme—Markov-chain-based clustering and farthest vector forwarding (MCBC–FVF). For intracluster energy optimization, a state transition probability model based on Markov chain (MC) is constructed for cluster head (CH) election. An energy-aware objective function with a spatial bias term is designed to reduce and balance energy consumption. For intercluster energy optimization, a farthest vector forwarding (FVF) mechanism is introduced to mitigate communication failure and low packet delivery ratio (PDR) caused by excessive distances between CHs. It also helps reduce redundant traffic and suppresses path inflation. Compared with MH-LEACH, LEACH-C, and MMRP, simulation results based on IEEE 802.15.4 demonstrate that the proposed MCBC–FVF scheme improves the first node death (FND) time by 38.68%, 22.62%, and 17.20%, respectively, while reducing intercluster average energy consumption by 32.73%, 18.41%, and 43.38%, respectively. These results indicate that MCBC–FVF not only significantly prolongs network lifetime but also provides a novel integration of probabilistic modeling and topologyaware forwarding, offering a practical and effective solution for energy-constrained WGNs.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 24","pages":"45050-45060"},"PeriodicalIF":4.3,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145729343","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-13DOI: 10.1109/JSEN.2025.3610164
Xueer Wang;Qi Teng
Presents corrections to the paper, (Corrections to “CSFO: A Category-Specific Flattening Optimization Method for Sensor-Based Long-Tailed Activity Recognition”).
提出了对论文的更正，（对“CSFO：基于传感器的长尾活动识别的特定类别平坦化优化方法”的更正）。
{"title":"Corrections to “CSFO: A Category-Specific Flattening Optimization Method for Sensor-Based Long-Tailed Activity Recognition”","authors":"Xueer Wang;Qi Teng","doi":"10.1109/JSEN.2025.3610164","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3610164","url":null,"abstract":"Presents corrections to the paper, (Corrections to “CSFO: A Category-Specific Flattening Optimization Method for Sensor-Based Long-Tailed Activity Recognition”).","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"25 22","pages":"42413-42415"},"PeriodicalIF":4.3,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11245647","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145500470","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-13DOI: 10.1109/JSEN.2025.3629733
You Tan;Kechen Song;Hongshu Chen;Yu Zhang;Yunhui Yan
The detection of internal surface defects in cold-drawn pipes is challenging. In recent years, as the production demands for cold-drawn steel pipes have steadily grown, there has been an urgent need for an efficient detection approach that balances accuracy and real-time performance in industrial environments. Although several existing deep learning-based methods have achieved high accuracy in surface defect detection, they often need substantial computational costs to extract rich feature representations, which inevitably slows down the inference process and leads to low detection efficiency. Moreover, internal defects of cold-drawn pipes typically exhibit challenges, which may further degrade the performance of existing models. To address these challenges, we propose a lightweight perception enhancement network (LPENet) to effectively balance efficiency and accuracy. Specifically, we introduce a progressive feature extraction (PFE) backbone that enhances contextual perception from local to global scales. Furthermore, we design amultiscale context enhancement (MCE) module to enrich the feature representation and a boundary-enhanced aggregation (BEA) module to strengthen fine-grained feature awareness. In addition, we propose a perception-guided fusion (PGF) strategy to facilitate interaction between shallow and deep features. We deploy LPENet in combination with a pipe internal surface detection (PISD) robot, achieving wireless and efficient defect detection in real-world steel pipe factories. In extensive experiments on the SSP2000 dataset, LPENet achieves the best balance between detection accuracy and efficiency. The source code is publicly available at https://github.com/VDT-2048/LPENet.
{"title":"A Lightweight Perception Enhancement Network for Real-Time and Accurate Internal Surface Defect Detection of Cold-Drawn Steel Pipes","authors":"You Tan;Kechen Song;Hongshu Chen;Yu Zhang;Yunhui Yan","doi":"10.1109/JSEN.2025.3629733","DOIUrl":"https://doi.org/10.1109/JSEN.2025.3629733","url":null,"abstract":"The detection of internal surface defects in cold-drawn pipes is challenging. In recent years, as the production demands for cold-drawn steel pipes have steadily grown, there has been an urgent need for an efficient detection approach that balances accuracy and real-time performance in industrial environments. Although several existing deep learning-based methods have achieved high accuracy in surface defect detection, they often need substantial computational costs to extract rich feature representations, which inevitably slows down the inference process and leads to low detection efficiency. Moreover, internal defects of cold-drawn pipes typically exhibit challenges, which may further degrade the performance of existing models. To address these challenges, we propose a lightweight perception enhancement network (LPENet) to effectively balance efficiency and accuracy. Specifically, we introduce a progressive feature extraction (PFE) backbone that enhances contextual perception from local to global scales. Furthermore, we design amultiscale context enhancement (MCE) module to enrich the feature representation and a boundary-enhanced aggregation (BEA) module to strengthen fine-grained feature awareness. In addition, we propose a perception-guided fusion (PGF) strategy to facilitate interaction between shallow and deep features. We deploy LPENet in combination with a pipe internal surface detection (PISD) robot, achieving wireless and efficient defect detection in real-world steel pipe factories. In extensive experiments on the SSP2000 dataset, LPENet achieves the best balance between detection accuracy and efficiency. The source code is publicly available at <uri>https://github.com/VDT-2048/LPENet</uri>.","PeriodicalId":447,"journal":{"name":"IEEE Sensors Journal","volume":"26 1","pages":"1383-1394"},"PeriodicalIF":4.3,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145852514","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：