Optical character recognition (OCR) has been widely applied in the computer vision, but the text recognition (TR) in the complex text scenes still faces the challenge of low TR accuracies. Therefore, a contrastive learning (CL)-based multi-scale attention (CLMSA)model for TR is proposed, aiming to improve the TR ability of TR model in the complex text scenes. Firstly, the modality transition module is improved by introducing the multi-scale attention mechanism to capture the global and local features in the image to enhance the ability of model to understand the complex text content. Secondly, a data augmentation-based invariant feature learning (DAIFL) strategy is studied to construct the enhancement domain by data augmentation and optimize the modality transition module based on the multiple-kernel maximum mean discrepancy (MK-MMD) to improve the robustness of the model. Thirdly, a character similarity-based contrastive loss (CSCL) is designed to construct the sample pairs by calculating the character similarities between different images to improve the TR effect of the model in the noisy scenes. Finally, a series of experiments are carried out to verify the effectiveness of the proposed model. The results reveal that the proposed CLMSA model can obtain excellent TRperformances in the complex text scenes.

Multi-sensor federated learning (MFL) can train a cross-sensor global model through distributed collaboration under the protection of data privacy, which has attracted attention in the field of fault diagnosis (FD). However, the sensor heterogeneity and unseen target domain (TD) pose key obstacles to the generalization of the MFL-based FD model, which seriously affect the performance of the global model in FD. Therefore, a novel sensor heterogeneous federated domain generalization (SHFDG) method based on prototype contrastive learning for intelligent FD is proposed. Firstly, a dual-level prototype contrastive (DLPC) strategy is proposed to mine the potential intrinsic associations in multi-source heterogeneous data through the contrast of prototypes at different levels, effectively breaking through the barrier of sensor heterogeneity and achieving the alignment of multi-sensor features. Secondly, a heterogeneous sensor hierarchical aggregation (HSHA) strategy is designed, which performs the hierarchical weighted aggregation on the cross-sensor global model to effectively address the aggregation bias caused by sensor heterogeneity and non-independent identically distributed (non-IID) data. Thirdly, a global prototype-guided classification (GPGC) strategy is designed, which dynamically weights and fuses the cosine similarities between the samples to be diagnosed and the global prototypes with the output of the fully connected (FC) layers to improve the stability and interpretability of fault classification. Finally, extensive experiments are carried out to verify the effectiveness of the proposed method. The results show that the method performs well in FD with sensor heterogeneity and unseen TD while ensuring data privacy.

The existing domain adaptation (DA) methods have been widely applied in cross-domain fault diagnosis (CDFD) of rotating machinery, but they typically assume that different domains share the same label space and rely on aggregating data from different domains for model training. However, under real industrial scenarios, the label spaces across different domains may be inconsistent due to complex and changeable working conditions, and the data from different domains may not be directly accessed due to data privacy. Therefore, a novel dual prototype-guided universal federated domain adaptation (DPUFDA) framework for CDFD is proposed. Firstly, a dual prototype-guided distribution alignment strategy is designed, where multiple class prototypes and a centroid prototype on each source client are constructed, aiming to promote the global domain and sub-domain alignments through the dual prototype contrastive learning. Secondly, an entropy-guided confidence discrepancy maximization strategy for the target client is designed to construct precise decision boundaries between the known and unknown fault classes, where the tailored discriminative confidence discrepancy entropy is adopted to enhance the classification confidence. Thirdly, a hierarchical hybrid classifier-based unknown-class rejection strategy is designed to accurately distinguish known fault classes and effectively reject unknown fault classes for the target client. Finally, a series of experiments are conducted to validate the effectiveness of the proposed framework, and the results confirm that the framework can effectively address the domain shift and label space shift while protecting data privacy, achieving excellent CDFD performance.

With the rapid development of autonomous driving, high accuracy and low latency are required for traffic sign detection, especially for small objects. However, achieving a balance between accuracy and speed is challenging due to excessive backbone computation, inadequate feature representation, and the design of separate tasks. FTC-YOLO is proposed to address these problems based on feature enhancement and task calibration. First, the network is enhanced by removing large detection layers that introduce unnecessary computation and by adding new layers for small objects so that detection performance is improved while efficiency is maintained. Then, an innovative StarNet backbone is utilized, employing star operations for rapid feature extraction. A Semantic-Detail Feature Enhancement Module (SFEM) is also designed, utilizing frequency domain transformations and global information to capture context from multiple directions, thereby enhancing the representation of small objects. Finally, a Joint Representation Calibration Detection Head (JRCDH) is proposed, which utilizes shared convolutions to enhance feature consistency across different representations. With deformable convolution (DCNv2) and channel attention, the alignment between classification and localization is improved. The bounding box loss function is further adjusted by combining Wise-IoU with the normalized Gaussian Wasserstein distance, making localization more stable and better at handling object shapes. On the TT100K dataset, FTC-YOLO achieves a mAP of 83.4

In actual industrial applications, due to the data privacy protection, the source domain (SD) data are often not directly accessible, while the fault classes that do not belong to the SD often appear on the target domain (TD). To address these challenges, a novel information entropy-guided source-free open-set fault diagnosis (IESOFD) approach is proposed. First, a graph convolutional network (GCN)-based source model is pretrained to extract robust features. The pretrained source model is transferred to the TD to obtain the target model, which is fine-tuned to achieve the knowledge transfer from SD to TD. Second, a multigroup partition pseudo-label (MGPPL) strategy is designed. It initially partitions the TD samples into multiple groups with the pseudo-labels, selects some samples from each group as the known and unknown class sample sets, respectively, and minimizes and maximizes the information entropy of the known and unknown class sample sets, respectively, aiming to separate the unknown fault classes in the TD and improve the reliability of the pseudo-labels. Third, a difficulty class learning (DCL) strategy is designed, which reduces the intraclass distances of the difficulty classes and uses different weights to learn each fault class by information entropy, so as to more effectively separate the known and unknown fault classes. Finally, extensive experiments are conducted to verify the effectiveness of the proposed approach, and the results confirm that the proposed approach can accurately identify the known and unknown fault classes in the TD while protecting the SD data privacy.

The existing domain generalization-based fault diagnosis (DGFD) methods typically rely on the assumption that the label spaces of the source domains (SDs) and the unseen target domain (TD) are consistent, and they aggregate data from different SDs for unified model training. However, in real-world industrial scenarios, the fault classes that are not covered in the SDs may appear in the TD, and the data privacy constraints prevent the direct aggregation of data from several SDs for model training. Therefore, a novel dual-task learning-driven and prototype-guided federated open-set domain generalization (FedOSDG) method for intelligent fault diagnosis (FD) is proposed. Firstly, a feature distribution-based dual-task learning (FDDTL) strategy is designed, which enables local models of different clients to learn domain-invariant features through feature compression based on the information bottleneck (IB) principle and adversarial learning. Secondly, a prototype-guided dynamic decision boundary construction (PGDDBC) strategy is proposed, which utilizes the local data from source clients to construct the local class prototypes and uses them as guidance to dynamically construct the optimal decision boundaries between different fault classes, aiming to effectively identify known and unknown fault classes. Thirdly, a prototype discrepancy-based dynamic aggregation (PDDA) strategy is designed, which aggregates local models by measuring the discrepancies between the local and global class prototypes to effectively alleviate the negative transfer problem caused by model aggregation. Finally, the extensive experiments demonstrate that the proposed method exhibits excellent cross-domain FD performance on the unseen TD under open-set scenarios while ensuring data privacy.

Defect detection in industrial components based on endoscopic images is a prevalent application in the industrial field. In practical scenarios, the images captured by endoscopes often exhibit characteristics such as low light and poor resolution, which lead to issues like low accuracy and high complexity in detection algorithms, posing challenges to existing industrial component inspection networks. To address these challenges, this paper proposes an Industrial Endoscope Defect Detection Network (IEDD-Net) based on YOLOv12n. First, a Dynamic ConvFormer-GLU Module (DCFG) is designed to enhance the detection capability for micro and small defects. Second, a pinwheel convolutional sampling layer (PCSL) is proposed, which enhances the extraction of defect features. In addition, a new detection head is redesigned to reduce the computational cost and model parameters while maintaining detection accuracy. The results demonstrate that, compared to the original model, IEDD-Net achieves improvements of 17.4

Federated transfer learning (FTL) has been widely applied in the rotating machinery fault diagnosis (RMFD). However, how to alleviate the impact of domain shift and avoid the negative transfer problem caused by model aggregation remain the key challenges that urgently need to be solved. Therefore, a prototype contrastive-based decentralized federated transfer learning (PCDFTL) approach for intelligent fault diagnosis is proposed. Firstly, a dual-phase adapter projection optimization (DPAPO) strategy is designed to ensure that the sample features can be mapped into a unified feature space by freezing the certain layers of the model while fine-tuning the other layers in the two phases, effectively achieving the domain alignment. Secondly, a prototype-based differentiated contrastive (PDC) loss is designed to mitigate the impact of domain shift by combining the multi-kernel maximum mean discrepancy (MK-MMD) and leveraging the target prototypes to guide the source models for domain alignment. Thirdly, by abandoning the idea of aggregating a global model, a multi-source dynamic weighted inference (MSDWI) strategy is studied to avoid the negative transfer problem by dynamically weighting and fusing the inference results of multiple source models on the target domain according to the distribution discrepancies between the source and target domains, the stability of domain adaptation (DA) process, and the certainty of the entropy of the inference results. Finally, extensive experiments are conducted to verify the effectiveness of the proposed approach. The results reveal that the approach can achieve excellent cross-domain RMFD performance while protecting data privacy. The code is available at https://github.com/HutAC/PCDFTL.

Adverse weather induces severe domain shifts in object detection for power line inspection via unmanned aerial vehicles (UAVs), compromising the detection of critical infrastructure components. Existing domain adaptation methods often assume a single degradation mechanism, inherently limiting their robustness across diverse, unconstrained environments. However, distinct weather conditions exhibit fundamentally different degradation physics—such as signal attenuation in haze and structured noise in rain—necessitating reliability-aware decoupling rather than uniform alignment, as such blind alignment may amplify pseudo-label noise during self-training. To address these challenges, we propose the Reliability-aware Teacher-Student Framework (R-TSF), a unified reliability-aware approach for robust object detection. Specifically, we first introduce a Weather-Aware Reliability Estimation (WARE) module to dynamically quantify local feature degradation without explicit weather classification. Guided by this reliability metric, we then deploy a Reliability-Guided Masked Autoencoder (R-GMAE) to suppress noise while recovering informative semantics, and employ a Reliability-Aware Variance Feedback Smoothing Thresholding (R-Aware VFST) strategy to stabilize pseudo-label generation across diverse degradation types. Comprehensive evaluations reveal that our approach attains mAP50 scores of 66.7

Open-set domain adaptation (OSDA) can address the realistic fault diagnosis (FD) scenario where the target domain contains previously unseen fault classes. However, most of the previous studies usually adopt coarse-grained separation approaches and ignore the important role of the closed-set model itself, which affects the FD performance. To address this limitation, a novel class prototype-guided OSDA (CPOSDA) method for mechanical FD (MFD) is proposed. First, a multisource adversarial pretraining approach is designed, which provides a high-quality closed-set model for constructing reliable cross-domain class prototypes. Second, a class prototype-guided progressive separation strategy is developed, which leverages the class prototypes for coarse-grained separation of target domain samples, followed by fine-grained training to refine the decision boundary between known and unknown classes. Third, a domain adaptation (DA) strategy with unknown-class rejection is designed to filter out samples from unknown classes during the distribution alignment process between the source and target domains, thereby reducing the risk of negative transfer. Finally, extensive experiments prove that the proposed method achieves excellent cross-domain FD performance in the open-set scenario and outperforms the comparison methods. The source code is available at https://github.com/NullIsFine/CPOSDA

The effectiveness of existing domain generalization-based fault diagnosis (DGFD) methods usually relies on the assumption that the label space of the source domain (SD) is consistent with that of the unseen target domain (TD). However, in actual industrial scenarios, the unknown fault classes that do not exist in the SDs may appear in the TD, resulting in the degradation of the diagnosis accuracies of DGFD methods on the unseen TD. Therefore, a novel open-set domain generalization (OSDG) approach via metalearning-based dual-level gradient alignment (MLDGA) for intelligent fault diagnosis (FD) is proposed. First, a metalearning optimization strategy with dual-level gradient alignment is designed to optimize the gradient update directions of the interdomain and interclass tasks simultaneously by gradient matching, so as to ensure that the decision boundaries are located in the optimal positions between each fault class. Second, an entropy-guided dynamic weighting strategy is designed to improve the discrimination ability and accuracy of the model in the multiclass fault classification tasks. Finally, a classification-clustering dual-guided open decision boundary construction strategy is designed to improve the recognition capability of unknown fault classes and the adaptability of the class decision boundaries in fault classification tasks. The experimental results confirm that the proposed approach can effectively identify both known and unknown fault classes.

Flexible job shop scheduling problem (FJSP) is a complex optimization problem in intelligent manufacturing and plays a key role in improving productivity, which is characterized by that each operation can be processed by multiple machines. Most current research into FJSP focuses on finding a higher-quality scheduling scheme in a shorter time. However, existing studies are hard to optimize the operation sequencing and machine assignment strategies simultaneously, which is critical for making the optimal scheduling decision. Therefore, a multi-agent-based graph reinforcement learning (MAGRL) method is proposed to effectively solve FJSP. Firstly, the FJSP is modeled into two Markov decision processes (MDPs), where the operation and machine agents are adopted to control the operation sequencing and machine assignment respectively. Secondly, to effectively predict the operation sequencing and machine assignment strategies, an encoder-double-decoder architecture is designed, including an improved graph attention network (IGAT)-based encoder, an operation strategy network-based decoder, and a machine strategy network-based decoder. Thirdly, an automatic entropy adjustment multi-agent proximal policy optimization (AEA-MAPPO) algorithm is proposed for effectively training the operation and machine strategy networks to optimize the operation sequencing and machine assignment strategies simultaneously. Finally, the effectiveness of MAGRL is verified through experimental comparisons with the classical scheduling rules and state-of-the-art methods to solve FJSP. The results achieved on the randomly generated FJSP instances and two common benchmarks indicate that MAGRL can consume less solution time to achieve higher solution quality in solving different-sized FJSP instances, and the overall performance of MAGRL is superior to that of the comparison methods.

Domain generalization-based fault diagnosis (DGFD) approaches do not require access to the target domain during model training, but they usually rely on numerous labeled source domain data. However, only few labeled source domain data can be obtained in actual diagnosis scenarios. Therefore, a novel hybrid data-driven domain generalization (DG) approach with dual-perspective feature fusion for intelligent fault diagnosis (FD) is proposed. Firstly, to solve the problem of scarce training samples in the source domains, the rolling bearing (RB) and the gear simulated vibration models are established to generate numerous labeled simulated vibration data, and the improved auxiliary classifier generative adversarial network (ACGAN) is used to effectively balance the simulated and real data. Secondly, a simulated and real data-driven DG network that fuses intra-domain invariant features and mutually-invariant features between domains (SRDGN-IM) is proposed, where the intra-domain invariant features are learned through distillation idea and the mutually-invariant features are learned through adversarial training, which can make the diagnosis model better learn the key generalization features from source domains to obtain more accurate diagnosis results. Finally, a series of DG experiments are conducted on the gearbox and bearing datasets, and the average FD accuracies of the proposed approach reach 87.45% and 89.10% respectively under different DG tasks.

The open shop scheduling problem (OSSP) involves complex processing constraints and a large number of job-machine combinations, which leads to an exponential increase in the solution space. For large-scale OSSP in real-world industrial productions, traditional methods struggle to provide satisfactory optimization results within a limited time. Therefore, an efficient method for solving large-scale OSSP through improved graph attention network based on link prediction (IGAT-LP) and Horovod-GPU is proposed. Firstly, an open shop scheduling (OSS) model based on IGAT-LP is designed to make full use of the feature information of operation nodes in OSSP. The model employs the graph attention network (GAT) structure to capture dependencies between tasks, learns global information through a multi-head attention mechanism, and predicts the optimal matching order between operations and machines. Secondly, a distributed parallelization method for the OSS model based on IGAT-LP is proposed. The distributed training capability of Horovod-GPU platform is fully utilized to expand the model training across multiple GPU nodes, significantly improving training efficiency. Finally, extensive experiments are conducted to analyze the effectiveness of the proposed method. The experimental results verify the superiority of the proposed method for solving large-scale OSSP instances. Moreover, the method significantly enhances the training performance of the OSS model based on IGAT-LP.

Wind power generation is an important approach to achieving clean energy and is associated with notable randomness and uncertainty. Wind power data exhibit strong periodicity and historical dependence, which are influenced by wind speed, temperature, and other environmental factors. Existing methods struggle to simultaneously capture complex time-space dependence and dynamic changes in wind power data, resulting in limited forecasting accuracy and generalization ability. To address these challenges in wind power forecasting (WPF), a method that can comprehensively account for the characteristics of spatiotemporal interaction and the long- and short-term dependencies in the data is urgently needed. Therefore, an efficient WPF method combining a multiscale graph convolutional network (MSGCN) and an extended long short-term memory (xLSTM) is proposed, referred to as the MSGCN-xLSTM. First, an MSGCN is designed to extract the spatial characteristics of each node, obtaining feature information across multiple timescales through multiscale graph convolution (MSGC) operations, where the characteristics of nodes and their neighbors are weighted and fused. Second, the global feature enhancement mechanism is employed to adjust the characteristics extracted over different timescales, capturing long-distance global dependencies. Third, the improved xLSTM is applied to model wind power time series characterized by strong periodicity and multivariate dependence, enhancing the capturing of long- and short-term dependencies. Finally, an enhanced global exploration-whale optimization algorithm (EGE-WOA) is used to adaptively optimize the hyperparameters of the model, thereby improving its adaptability to complex wind power data. Furthermore, extensive experiments conducted on historical datasets from real wind farms demonstrate that the proposed method effectively captures the complex interactions and dynamic changes in spatiotemporal characteristics of wind power data. The proposed method not only surpasses comparison models in forecasting accuracy and generalization on datasets from different wind farms but also excels in training efficiency and convergence.

In actual industrial production, the working conditions of rotating machinery are complex and changeable, and the health-state monitoring data are increasingly large and difficult to label, which will seriously restrict the accuracy and efficiency of the cross-domain fault diagnosis (CDFD) of rotating machinery. Therefore, an efficient multi-source domain deep transfer learning (MDDTL) method for CDFD of rotating machinery is proposed. First, an MDDTL model is constructed to improve the accuracy of CDFD. In the model, a dual-phase domain alignment strategy is designed, which considers the alignment of feature distributions between each source and target domain pair in the feature space and that of the prediction probabilities between domain-specific fault classifiers in the output space. The fault prediction results from multiple different fault classifiers are merged dynamically by the proposed imbalanced adaptive prediction strategy. Secondly, a data-parallel distributed training scheme for the MDDTL model is proposed. Based on the idea of data parallelism, the distributed parallel training of the MDDTL model is performed with a Horovod-graphics processing unit platform, and the parameters are synchronously updated with the bandwidth-optimal Ring-AllReduce architecture. Under the premise of ensuring the accuracy of FD, the training time of the MDDTL model is significantly reduced. Finally, extensive experiments are conducted to verify the effectiveness of the proposed MDDTL method. The results demonstrate that the proposed method not only effectively improves the accuracy of CDFD of rotating machinery but also significantly improves the training efficiency of the MDDTL model. After adopting the proposed method, the diagnosis accuracies achieved under two different cross-working condition scenarios reach 97.09% and 97.87% respectively, and the model training time is reduced by 73.62% when facing a large-scale rotating machinery training set.

In practical industrial applications, it is crucial to train a robust fault diagnosis (FD) model that can quickly adapt to new working conditions or fault modes using a few labeled fault samples. Therefore, a novel convolutional multi-head self-attention network-based meta-transfer learning approach (CMS-MTL) for few-shot fault diagnosis (FSFD) is proposed. Firstly, a convolutional multi-head self-attention network (CMHSAN) is designed, which ingeniously combines the multi-head self-attention (MHSA) blocks and convolution blocks. The local and global feature information of the input time–frequency images are fully considered through the mutual cooperation of MHSA and convolution, so as to fully extract the discriminative features among various fault classes. Secondly, a three-stage CMHSAN-based meta-transfer learning (MTL) scheme is proposed, which provides a good initialization state for the meta-training of the CMHSAN model through the pre-training stage, updates the pre-trained model with the scaling and shifting parameters in the meta-training stage, and fine-tunes the updated model in the meta-testing stage, so as to quickly adapt to new FSFD tasks from the target domain. Thirdly, aiming at the fault classes that are difficult to be diagnosed during meta-training, a meta-task re-training (MTRT) strategy is designed to learn more valuable transferable knowledge in the meta-training stage, thereby improving the adaptability of the CMHSAN model to hard FSFD tasks. Finally, extensive experiments are conducted under different FSFD scenarios to verify the effectiveness of the proposed approach. The results prove that the approach can quickly adapt to new FSFD tasks through the learned meta-knowledge and achieve high diagnosis accuracies.

Heterogeneous platforms composed of multiple different types of computing devices (such as CPUs, GPUs, and Intel MICs) have been widely used recently. However, most of parallel applications developed in such a heterogeneous platform usually only utilize a certain kind of computing device due to the lack of easy-to-use heterogeneous cooperative parallel programming models. To reduce the difficulty of heterogeneous cooperative parallel programming, a directive-based heterogeneous cooperative parallel programming framework called HeteroPP is proposed. HeteroPP provides an easier way for programmers to fully exploit multiple different types of computing devices to concurrently and cooperatively perform data-parallel applications on heterogeneous platforms. An extension to OpenMP directives and clauses is proposed to make it possible for programmers to easily offload a data-parallel compute kernel to multiple different types of computing devices. A source-to-source compiler is designed to help programmers to automatically generate multiple device-specific compute kernels that can be concurrently and cooperatively performed on heterogeneous platforms. Many experiments are conducted with 12 typical data-parallel applications implemented with HeteroPP on a heterogeneous CPU-GPU-MIC platform. The results show that HeteroPP not only greatly simplifies the heterogeneous cooperative parallel programming, but also can fully utilize the CPUs, GPU, and MIC to efficiently perform these applications.

Due to the dynamic changes of manufacturing environments, heuristic scheduling rules are unstable in dynamic scheduling. Although meta-heuristic methods provide the best scheduling quality, their solution efficiency is limited by the scale of the problem. Therefore, a novel method for solving the dynamic flexible job-shop scheduling problem (DFJSP) via diffusion-based transformer (DIFFormer) and deep reinforcement learning (D-DRL) is proposed. Firstly, the DFJSP is modeled as a Markov decision process, where the state space is constructed in the form of the heterogeneous graph and the reward function is designed to minimize the makespan and maximize the machine utilization rate. Secondly, DIFFormer is used to encode the operation and machine nodes to better capture the complex dependencies between nodes, which can effectively improve the representation ability of the model. Thirdly, a selective rescheduling strategy is designed for dynamic events to enhance the solution quality of DFJSP. Fourthly, the twin delayed deep deterministic policy gradient (TD3) algorithm is adopted for training an efficient scheduling model. Finally, the effectiveness of the proposed D-DRL is validated through a series of experiments. The results indicate that D-DRL achieves better solution quality and higher solution efficiency when solving DFJSP instances.

In the domain of medical image segmentation, the nnUNet framework is highly respected for its excellent performance and wide range of applications. However, the inherent bias of locality and weight sharing introduced by the continuous convolutional operations currently used limits the network’s performance in modeling long-term dependencies. Furthermore, in the process of implementing residual links, certain limitations are encountered due to the substantial semantic discrepancy between the encoder’s output feature maps and the decoder’s. These limitations are seen in the direct application of skip connections for feature fusion and gradient propagation, which are known to impact the model’s convergence speed and overall performance. In this paper, a novel framework is presented, namely Multi-Attention nnUNet (MulA-nnUNet), which utilizes nnUNet as the foundational network structure and integrates two key attention mechanisms: large kernel convolutional attention (LKA) and pixel attention (PA). LKA is embedded within the deep encoder, maintaining the effectiveness of shallow feature extraction and enhancing the deep neural networks’ ability to understand long-range spatial dependencies. At the same time, the semantic distinction between the encoder and decoder’s output map of features is decreased by the PA module, which helps to improve the effect of skip connection feature fusion. The complexity of the model is reduced by replacing the standard convolutions in the encoder and decoder layers with depthwise separable convolutions (DS), which have fewer parameters. The effectiveness of the proposed framework is confirmed by a set of ablation experiments and comparison experiments with current state-of-the-art models on the computed tomography (CT) subset of the multimodal abdominal multi-organ segmentation dataset (AMOS), which includes 500 CT scans, with 350 scans for training, 75 for validation, and 75 for testing. MulA-nnUNet shows improvements of 1.1% in mean dice similarity coefficient (mDSC) and 1.52% in mean intersection over union (mIoU), while the baseline model requires 5 times the floating point operations (FLOPs) and over 7 times the parameters (Params). Additionally, it demonstrates superior accuracy in segmenting organs such as the liver, stomach, aorta, and pancreas, thereby enhancing the accuracy of 3D abdominal multi-organ image segmentation.