Abstract:In emergency communication scenarios, unmanned aerial vehicles (UAVs) serve as aerial data collection platforms that can be rapidly deployed to disaster-stricken areas such as those affected by earthquakes, floods, wildfires, mining accidents and battlefield environments. UAVs are capable of collecting data from wireless IoT devices and transmitting it to the command center, thereby improving the efficiency of rescue decision-making. In post-disaster scenarios, data transmission tasks impose higher requirements on both communication efficiency and data completeness. Meanwhile, the limited energy supply of UAVs makes it challenging to collect large volumes of data efficiently while ensuring the complete transmission of vital information. To address these issues, this paper investigates a wireless communication system assisted by a single UAV, which adopts a multi-user uplink communication mode and a fly-hover-communicate data collection pattern. A joint optimization problem is formulated for device association, UAV hovering location and bandwidth allocation, aiming to maximize the UAV′s coverage utility while minimizing its total energy consumption. First, to enhance the UAV′s coverage utility, a particle swarm optimization (PSO) algorithm initialized with K-means clustering is employed. Then, to minimize energy consumption, we propose a PSO-based two-stage optimization framework that alternately optimizes hovering positions and bandwidth allocation. In particular, a PSO variant incorporating Gaussian perturbation and differential mechanisms is designed for hovering position refinement. Simulation results demonstrate that the proposed method effectively improves both coverage utility and energy efficiency. The coverage utility increased by 13.15% compared to the K-means algorithm, while energy consumption was reduced by 18.24% compared to the approach that only optimizes hovering locations and lower than the scheme where hovering energy and flight energy are optimized separately.

Abstract:There are many sensors for civil aviation engine health monitoring. The proper choice of sensors will directly affect the prediction effect of engine remaining useful life. A sensor selection method based on optimal feature selection is proposed and Informer algorithm is used to predict the remaining useful life, which improves the prediction accuracy. Firstly, the differential clustering algorithm is used to classify the real flight conditions, and the health factors are constructed from the degradation mechanism of civil aviation engine, and the regression tree model is established with the data of cruise stage to select important sensors. Finally, the remaining useful life of civil aviation engine is predicted based on Informer algorithm. Using NASA′s newly released civil aviation engine degradation database under real flight conditions, the experimental results show that the root mean square error of prediction results decreases by 14% and the average scoring function decreases by 29% compared with no sensor selection. Compared with the traditional selection method based on sensor degradation trend or sensor data difference, the root mean square error decreases by 10% and 8% respectively, and the average scoring function decreases by 48% and 27% respectively. Compared with CatBoost, LightGBM, XGBoost, BiLSTM and Transformer algorithms, the accuracy of the proposed remaining life prediction method is improved by 36%, 24%, 14%, 6% and 5%, respectively.

Abstract:A fault diagnosis framework for chemical processes is proposed, which combines K-means synthetic minority oversampling techniquewith conditional adversarial domain adaptation to address issues such as feature coupling caused by the temporal dependence of multivariate sensing data, data distribution shift caused by changes in operating conditions, and imbalanced sample data in chemical processes. Firstly, the original one-dimensional data is converted into multiple two-dimensional time window data using time window segmentation technology. Within these windows, the Kmeans SMOTE method is used to expand the minority class fault samples. The expanded samples can retain the complete temporal fault features, and this algorithm can also reduce the number of generated noise samples; then, domain adaptation techniques are used to align the feature distributions of the source domain and the target domain, reducing the distribution differences between the two and enabling the fault diagnosis model trained on the source domain to effectively identify fault categories under new operating conditions; finally, diagnostic experiments were conducted using fault data from the Tennessee Eastman process, and the effectiveness of the proposed method was validated by comparing its diagnostic rates with models such as CDAN, DANN, and JDA.

Abstract:Aiming at the challenges of Doppler frequency shift, data bit transition, and wide spreading factor adaptability faced by DSSS signal acquisition technology in satellite-ground TT&C systems, this paper proposes a joint acquisition algorithm based on truncated PN code segmentation. By establishing a truncated PN code segmented parallel correlation architecture combined with an N-segment time-domain aggregation strategy, and implementing two-dimensional joint search through Fast Fourier Transform (FFT)-based frequency offset estimation, the algorithm effectively suppresses correlation peak attenuation caused by Doppler frequency shift (±800 kHz) and bit transitions. In terms of hardware architecture optimization, we analyze the sampling distribution patterns between truncated PN code segments and reconstruct parallel correlator resource allocation, reducing multiplier resource consumption by 73% compared to traditional dual-segment schemes. Simulation results demonstrate that under low spreading factor scenarios (SF=12), when truncation parameter N=32 is selected, the correlation peak-to-noise difference increases ninefold at chip SNR=-18 dB, achieving 89% detection probability (false alarm probability ≤2.5×10-5). FPGA implementation under identical conditions shows stable detection probability exceeding 85%. Featuring dynamically adjustable truncation parameters, this solution overcomes limitations of conventional fixed architectures in high-dynamic rate-switching scenarios, providing an engineering solution with high performance and low complexity for miniaturized spaceborne equipment and high-dynamic weak signal acquisition.

Abstract:Time series data pose unique challenges for forecasting due to their complex long- and short-term patterns and multi-period characteristics. Traditional fixed-scale patching methods are difficult to effectively capture multi-period information, while periodicity and trend changes further increase the modeling difficulty, affecting forecast accuracy and interpretability. Based on the above problems, this paper proposes a multi-periodic model MDTDNet based on dual time-dependent learning. The model firstly acquires multi-period information adaptively by Fourier transform; then for each period, combined with the seasonal trend enhancement module, it improves the semantic expression of the subsequence through period patch design, frequency domain seasonal enhancement and time domain trend enhancement. A dual time-dependence module is introduced to realize feature extraction and fusion by capturing different time-dependent patterns inter- and intra-patchs by means of a long-term change extractor and a local fluctuation extractor, respectively. The experimental results show that the experimental results of the models in all six datasets outperform the current optimal model, PatchTST, with an average decrease of 9% in the mean square error (MSE) on ETTh1 dataset, with a maximum decrease of 10.14%.

Abstract:In view of the problems such as poor adaptability to complex working conditions, insufficient nonlinear error compensation accuracy and low level of instrument intelligence in high oxygen and trace oxygen measurement in the field of gas detection, this paper designs a gas detection of BP neural network (MAPSO-BP) optimized by lightweight modified adaptive particle swarm algorithm. Instrument, the system builds a multi-sensor embedded platform to realize synchronous acquisition and fusion compensation of multi-parameters including temperature, pressure, flow and concentration, uses a microcontroller unit to run MAPSO-BP network in real time for nonlinear error correction, and develops an embedded human-computer interaction system based on Qt, supporting network communication, data storage, real-time alarm and cloud data synchronization functions enhance the intelligence level of the instrument. The system prototype designed in this paper is tested for system stability, anti-interference ability test and comparative experiments with the existing error compensation model. The results show that the error compensation method proposed in this paper and the designed system prototype are compared with the current mainstream error compensation method. The absolute error average of high oxygen and micro-oxygen measurements are respectively reduce by 20% and 25%; effectively solve the problem of low measurement accuracy of sensors under complex working conditions, and provide a feasible solution for the precision and low cost of gas sensors.

Abstract:Aiming at the recognition failure of traditional 2D code detection methods for assisted navigation in complex industrial, logistics and transport scenarios, this study proposes an improved and lightweight YOLOv8-AR model to further enhance the recognition efficiency by using AR codes that are mature and can provide relative positional information. In terms of the network model, the backbone introduces an ultra-strong lightweight StarNet network to reduce the algorithmic computation of target detection; the C2f-EMSC module is optimised and constructed in the neck network to enhance the extraction of AR code features in complex environments and reduce the computational load at the same time; moreover, a lightweight detail-enhanced shared convolutional detection head LSDECD-H is proposed to improve the detail feature expression capability, so as to improve the detection accuracy of small targets and multi-targets. The experimental results show that the parametric and computational quantities of the model are 1.46M and 4.7GFLOPs, which are only 51% and 42% of the baseline, and the mAP is as high as 0.962 with high robustness in the case that the frame rate meets the real-time detection. It can quickly determine the position before decoding, and improve the recognition effect to achieve precise positioning, making it suitable for application scenarios like 2D code road sign navigation.

Abstract:Ground current in high-voltage cable systems serves as a critical indicator for ensuring operational safety and stability. Accurate ground current prediction is crucial for fault prevention and enhancing grid reliability. To address the limitations of traditional time-series prediction models in terms of prediction accuracy and computational efficiency, this paper proposes a ground current prediction model based on the Mamba architecture, referred to as the Bi-EMamba model. Through a spatiotemporal dependency encoder, the model effectively captures long-term dependencies and spatial correlations in multivariate time series while maintaining high memory efficiency. To address the challenge of non-stationary data, the model incorporates Reversible Instance Normalization for data normalization and employs hyperparameter optimization to further improve prediction accuracy and generalization capability. Experimental results based on a dataset from a high-voltage cable line in Beijing demonstrate that Bi-EMamba outperforms existing benchmark models across various prediction horizons. Notably, in long-term forecasting scenarios, it exhibits superior generalization and computational efficiency. Compared to the current state-of-the-art model, iTransformer, Bi-EMamba achieves a 6.52% reduction in Mean Squared Error, a 3.21% reduction in Mean Absolute Error, and a 29.49% reduction in memory usage.

Abstract:Anomaly detection is an important task in modern industrial manufacturing. Due to the scarcity of abnormal samples, unsupervised detection that only requires normal sample training has attracted widespread attention. Among them, reconstruction based detection has been widely applied due to its concise and universal framework. However, existing algorithms are mostly based on image reconstruction, thus the discrimination between abnormal and normal regions is insufficient. At the same time, due to the strong uncertainty of abnormal positions and sizes in industrial images, existing algorithms cannot capture the overall structural features of samples well. This article proposed an industrial image anomaly detection algorithm based on feature reconstruction to address the above issues. Firstly, the use of pre trained models to extract multi-scale features as reconstruction objects avoids the situation where pixel space reconstruction has insufficient ability to distinguish anomalies; secondly, a global feature extraction module was designed to enhance the perception ability of the reconstruction model towards global features; finally, design a feature recombination strategy to jointly train the reconstruction model, in order to further enhance the model′s understanding of the overall structure of the samples and improve the reconstruction effect. A large number of experiments conducted on MVTec AD have shown that the proposed algorithm achieves an AUROC score of 98.7% in sample level anomaly detection and 98.3% in pixel level anomaly localization, both of which have reached state-of-the-art performance.

Abstract:The demand of modern warfare has propelled the application of multi-UAV collaboration in the military field. To address the problem of trajectory planning for multiple UAVs in a multi-threat mountainous environment with radar, artillery, and other threats, an improved Crested Porcupine Optimizer (CPO) algorithm, namely BCPO, is proposed.To tackle the issue of population diversity, the algorithm incorporates an initialization method combining opposition-based learning and good-point set initialization, which enhances the algorithm′s traversal capability. For the development phase of the CPO algorithm, a spiral search strategy based on adaptive small perturbations is introduced to further boost the global search performance. In the exploration phase of the CPO algorithm, a mutation triangle walk strategy based on the optimal random position is added to improve the local convergence efficiency. Additionally, a L-vy flight strategy with dynamic factors is proposed to help the algorithm achieve a better balance between global search and local optimization.Simulations on the CEC2017 test functions demonstrate that the BCPO algorithm has excellent convergence speed and accuracy. In a simulated mountainous environment, the BCPO algorithm shows an average performance improvement of 8.834%, 5.776%, and 21.828% compared to the CPO, GWO and WOA, respectively. Moreover, the stability of the algorithm is significantly enhanced. This method has good application value in solving multi-UAV trajectory planning problems in complex scenarios.

Abstract:The artificial lemming algorithm is a newly proposed metaheuristic method that simulates four distinct behaviors of lemmings to effectively explore complex search spaces. However, it still suffers from premature convergence, limited exploration, lack of robustness, and susceptibility to local optima. To address these limitations, a multi-strategy improved artificial lemming algorithm is proposed. First, the Halton sequence is employed to generate a uniformly distributed initial population, enhancing global search capability. Second, an elite pool strategy combined with inertia weights is introduced to reduce excessive reliance on the best individuals and to improve the population′s ability to jump across the search space, thereby suppressing premature convergence. Finally, a nonlinear weighted golden sine strategy, combined with foraging behavior, is incorporated in the later stages of iteration to enhance the precision and stability of local search. To verify the performance of the improved algorithm, experiments are conducted on the CEC2017 benchmark function set, and statistical analysis is performed using the Wilcoxon rank-sum test. Experimental results show that the improved algorithm outperforms five comparative algorithms in terms of convergence speed, optimization accuracy, and stability. Compared to the original algorithm, the improved version achieves an average error reduction of 27.36% and a reduction of 36.99% in the average standard deviation. In three engineering optimization problems, the improved algorithm obtains the minimum objective function values in all cases, demonstrating better applicability and superiority over the comparative methods.

Abstract:To meet the practical demands of autonomous ships for self-identifying navigation scenarios at sea, an adaptive recognition method for navigation scenarios in port waters based on electronic nautical charts is proposed. Firstly, by systematically analyzing the navigation characteristics of port waters, the ship navigation process is divided into eight scenarios: entering the port, leaving the port, navigating in the channel, entering and leaving the anchorage, anchoring, berthing, unberthing and mooring in the port. Secondly, a scene determination rule is established based on the characteristics of objects and their relative positions, and an adaptive recognition model integrating geometric relationships and dynamic parameters is constructed. Finally, the proposed method is verified based on the AIS historical trajectory data of 100 ships entering and leaving the waters of Yantai Port. The results show that the precision rate of this method for port navigation scenarios reaches 95.6%, and the delay is reduced to 12 ms. It can provide real-time necessary navigation scenario information and high-precision navigation environment perception support for the autonomous navigation system of ships sailing in the port area.

Abstract:To address the limitations of the CAM++ model in feature extraction and recognition performance under complex acoustic conditions, this paper proposes TF-DCAM, a speaker verification model integrating dilated convolution and temporal-frequency multi-scale attention mechanisms. The model enhances feature representation through dilated residual convolution and a time-frequency adaptive refocusing unit to suppress redundant information. A temporal-frequency multi-scale attention module is introduced to improve sensitivity to key information via channel attention and cross-dimensional interaction. An adaptive masking temporal convolution module is further incorporated to model long-term dependencies effectively. Finally, a combination of contrastive loss functions is applied to jointly optimize the speaker embedding space. Experiments conducted on the CN-Celeb dataset show that TF-DCAM reduces EER and minDCF by 14.98% and 10.98% respectively, compared with the baseline. The model also demonstrates strong cross-lingual generalization on the VoxCeleb1 dataset. Results indicate that the proposed method significantly improves speaker verification performance and robustness while maintaining model efficiency.

Abstract:This paper proposed an Event-triggered Curriculum DDPG algorithm to improve the efficiency and accuracy of dynamic target tracking for UAVs. The algorithm combined Deep Deterministic Policy Gradient （DDPG） and YOLO object detection technology. It introduced an event-triggered mechanism to dynamically adjust the policy update frequency, enhancing decision-making efficiency. Additionally, it incorporated curriculum learning to create a staged training framework, gradually improving the UAV′s adaptability to complex tasks. Experimental results showed that the ETC-DDPG algorithm effectively improved the tracking efficiency of dynamic target tracking task and the stability of training process, and reduced the demand for computing resources, achieving a success rate of 93.357%. Compared with the original-DDPG algorithm and ETC-TD3 algorithm, the success rate is improved by 56.175% and 37.1% respectively. The collaborative effect of the event-triggered mechanism and curriculum learning was verified by ablation experiment, providing a reference for autonomous task execution in UAVs.

Abstract:This paper, based on the principle of eddy electromagnetic induction, specifically the eddy current method and the magnetic method, combined with finite element simulation technology, simulates and analyzes the coil impedance and magnetic induction strength of the substrate surface, and verifies the feasibility of the inspection method. According to the study, the galvanized layer and coating thickness of the substrate surface have a significant effect on the coil impedance and magnetic induction strength. Based on this, this paper designs and develops the measurement circuit, algorithm and detection equipment for the thickness of galvanized layer and coating on the surface of ferromagnetic substrate. The experimental results show that the error of this equipment is less than 1% in the measurement of coating thickness on aluminum base surface, less than 6% in the measurement of coating thickness on iron base surface, and less than 6% in the measurement of insulator cap thickness. The coating layer detector developed in this paper can measure the thickness of the galvanized layer and the coating layer simultaneously, efficiently and accurately.

Abstract:NdFeB (neodymium-iron-boron) permanent magnetic materials have been widely applied in modern industry and electronics due to their exceptionally high magnetic energy product and coercivity. However, in practical production, the compaction process— a critical stage in NdFeB manufacturing— still relies primarily on operator experience for setting process parameters. Variations in operator expertise and the inherent complexity of the production process often lead to unstable parameter settings, which adversely affect product quality and result in resource wastage. To accurately predict the process parameters during the powder compaction stage, this study proposes a Dynamic Layered Adjustment CatBoost (DLA-CatBoost) multi-output prediction model. Furthermore, an innovative hybrid optimization strategy, PSO-DSS-NSGA-III, which integrates particle swarm optimization to guide dynamic search space adjustment, is introduced to achieve multi-objective cooperative optimization of the model′s hyperparameters. Experimental results demonstrate that the DLA-CatBoost model optimized with the PSO-DSS-NSGA-III strategy exhibits excellent performance in multi-output prediction tasks, with a root mean square error (RMSE) ranging from 0.5 to 0.9, a mean absolute error (MAE) between 0.2 and 0.5, and a coefficient of determination (R2) between 0.96 and 0.99, thereby demonstrating its superior predictive capability and establishing it as an effective new approach for optimizing the process parameters in NdFeB compaction.

Abstract:Addressing the issue of false negatives and positives in non-motor vehicle irregular driving behavior detection with the current detection algorithm, an improved target detection algorithm, YOLO-CSSM, was proposed based on YOLOv8. The Backbone and Neck were enhanced with an SPD-Conv network module, which improved the model′s ability to learn from small targets and extract features under complex backgrounds. Subsequently, DCNv2 and SegNext Attention modules were integrated into the Backbone and Neck networks, respectively, to emphasize important feature information of non-motor vehicles and drivers, enhancing the model′s feature fusion capability. The MPDIoU was improved using the concept of the WIoU loss function, replacing the original CIoU loss function with Wise-MPDIoU to mitigate the imbalance between positive and negative samples. Validated on a self-built dataset of non-motor vehicle irregular driving behaviors, the improved YOLOv8 algorithm demonstrated precision, recall and mean average precision (mAP@0.5) of 89.4%, 90.0% and 93.6%, respectively, showing improvements of 3.3%, 5.4% and 4.5% over the traditional YOLOv8 algorithm, achieving better detection accuracy and effectiveness.And Based on the non-motorized vehicle violation detection algorithm, a non-motorized vehicle violation recognition and detection system was designed and developed using PyQT5.

Abstract:This paper proposes a robust recognition method based on dual attention calibration to address the issues of insufficient multi-dimensional dynamic collaboration and fine-grained suppression in the attention mechanism under mask occlusion. The method dynamically calibrates the occlusion area in both channel and spatial dimensions. The channel dimension is based on global statistics to suppress abnormal responses of polluted channels, while the spatial dimension locates occluded areas and weakens their gradient propagation, achieving dynamic calibration from coarse-grained screening to fine-grained enhancement. On this basis, the weighted cross entropy loss and triplet loss are used to further guide the model to focus on the feature expression of locally unobstructed areas, thereby expanding the inter class feature distance interval. The experimental results show that the dual attention calibration mechanism proposed in this paper, through the synergistic effect of channel dimension feature screening and spatial dimension region enhancement, has improved accuracy by 6% and 7.2% respectively compared to the ArcFace algorithm in mask scenes of LFW and AgeDB-30, and by 7.3% on the real occlusion dataset MAFA dataset, verifying its recognition robustness in complex occlusion scenes.

Abstract:In response to the challenges posed by small target volumes and complex backgrounds in aerial remote sensing images, a lightweight object detection algorithm named ELS-RTDETR, based on enhancements to RT-DETR, has been proposed. This algorithm introduces and utilizes a new backbone network called LOB-Vovnet, which is an improved version based on the Vovnet network, to replace the original backbone network.Within the LOB-Vovnet architecture, a novel feature enhancement module named LRFF (Lightweight receptive field focus) has been designed to enhance the detection accuracy of small targets. To address complex background interference, an attention mechanism called SE (Squeeze-and-Excitation) based on adaptive channel extraction has been introduced.To strike a balance between model accuracy and size, LOB-Vovnet replaces some convolutions with depthwise separable convolutions. Extensive ablation experiments have been conducted to readjust the depth and width of the backbone network. In the AIFI section, a Cascaded Group Attention （CGA） mechanism has been introduced to effectively reduce computational redundancy in multi-head attention mechanisms.Regarding datasets, the RSOD dataset and NWPU VHR-10 dataset have been merged. Additionally, offline data augmentation techniques such as affine transformations and camera noise have been applied to the original data to make the training dataset more closely aligned with real-world applications.Experimental results indicate that the improved ELS-RTDETR model has shown a 2.7% increase in mAP@50 compared to the original model, with a reduction in model parameters by 32.9%. It has demonstrated good detection performance for challenging targets. Further validation of the enhanced method has been conducted on the SIMD dataset to verify its effectiveness.

Abstract:Existing pedestrian re-identification algorithms heavily rely on convolutional neural networks as the backbone, which often leads to an overemphasis on regions with prominent features while neglecting broader foreground features. This results in insufficiently rich global feature representations and inadequate attention to subtle discriminative features. To address these issues, we propose a feature-enhanced pedestrian ReID algorithm. The global branch utilizes position encoding and a multi-layer, multi-head attention structure to better leverage spatial context information, enhance the understanding of relative spatial positions, and effectively capture spatial structural information, thereby improving feature representation and global feature extraction capability. The local branch optimizes spatial attention using feature matrices associated with spatial vectors, enabling the capture of more compact general appearance features. Furthermore, by modeling the relationships between different channels, it strengthens feature expression in the channel dimension, highlighting distinctive features and improving the attention to discriminative characteristics. Finally, the model is trained using softmax loss, triplet loss, and center loss on the Market-1501 and DukeMTMC-ReID datasets. Experimental results demonstrate the effectiveness and superior performance of the proposed algorithm.

Abstract:To address the insufficient intelligent detection of surface damage on general aviation aircraft skins, an improved YOLOv11n-based detection algorithm is proposed. Firstly, the Adown downsampling mechanism is replaced to construct a multi-scale feature fusion architecture, achieving dynamic compression of redundant information through cross-level feature interaction and lightweight kernel design, thereby reducing model parameters and computational complexity. Secondly, a DySample dynamic upsampling strategy is designed, enhancing the model′s generalization across different scenarios via variable convolutional kernel deformation perception and multi-task gradient collaborative optimization. Furthermore, the FASSHead feature aggregation module is introduced, improving the algorithm′s recognition capability for complex damage areas through progressive semantic fusion and edge-aware constraints. Finally, a P2 small object detection layer is added, embedding high-resolution detection branches in shallow networks to enhance the capture of small objects and detailed damages. The improved algorithm was validated using a self-built dataset of general aviation skin surface damages. Results show that the precision reached 87.4%, recall reached 80.4%, and mAP attained 86.6%. Compared with the baseline model YOLOv11n, these metrics improved by 2.0%, 9.4% and 6.7% respectively, significantly enhancing the detection performance of skin surface damage and laying a theoretical foundation for an intelligent detection and maintenance system for general aviation aircraft.

Abstract:Aiming at the limitation of face detection accuracy degradation when facing light changes or complex background in driver fatigue detection methods, an improved MTCNN network is proposed. By optimising the MTCNN network, the coordinate attention mechanism and batch normalisation algorithm are introduced in all three sub-networks to improve the model′s localisation accuracy of the driver′s face, enhance the convergence speed and stability of the network, and enhance the suppression of overfitting. The experimental results show that the accuracy of the improved MTCNN model on the fatigue driving dataset reaches 98.78%, which is 2.43% higher than that of the original model, and the number of parameters of the model is only 0.5 M, which has good face detection accuracy and deployability. In addition, combining the improved MTCNN model with the PFLD model, a reasonable fatigue parameter threshold is set based on the experiments, and a more accurate fatigue driving detection is achieved.

Abstract:Crack detection is crucial in the maintenance of civil infrastructure. The many drawbacks of traditional manual visual inspection methods have led to the continuous development of crack detection methods. However, existing crack detection techniques face the challenges of complex backgrounds and feature diversity interference, and the high computational resource requirements. This study exploits the potential of Mamba for visual tasks and proposes an UltraLight CrackNet, which consists of a parallel lightweight visual Mamba block for efficiently modelling long-distance dependencies and extracting deep semantic features, a multi-scale residual visual state space block for enhanced multi-scale feature representation, and an enhanced semantics and detail infusion module for optimising skip connections within the encoder-decoder architecture. The experimental results show that our method requires only 0.13 M parameters and 1.96 G FLOPs, and achieves the optimal performance on DeepCrack and Crack500 datasets with ultra-lightweight model design, with the mean intersection over union (mIoU) of 87.85% and 77.92%, respectively, and obtains comparable results on SteelCrack dataset, and the number of parameters is 87.85% lower than that of the model with the smallest number of parameters among the available comparison models.