Abstract:A path planning method based on an improved synchronous bidirectional A-star algorithm and grey wolf optimization algorithm is proposed for the multiobjective cruising path planning problem of unmanned boats. Firstly, the traditional A-star algorithm has been improved by introducing a synchronous bidirectional search strategy and dynamic weight adjustment, reducing path redundancy points and algorithm computation time. Then, the cruise path planning problem is transformed into a classic traveling salesman problem and solved using an improved grey wolf optimization algorithm to obtain the optimal cruise path. The experimental results show that the method proposed in this paper is superior to traditional methods in terms of total distance, number of turns, and computation time in path planning. It can effectively improve the cruising efficiency and safety of unmanned boats and provide a reliable solution for multi-target point cruising tasks of unmanned boats.

Abstract:The high cost of collecting and processing high-quality motor fault data samples has resulted in the collection of newly unlabeled data samples. Domain adaptation has emerged as a promising approach to process and recognize new unlabeled data with the help of existing data. This has led to a surge of interest in domain adaptation in the field of fault diagnosis. In the field of electric machine fault diagnosis based on domain adaptation, two issues require attention. A conflict arises in the gradients of multiple tasks when employing the conventional domain adaptation framework. And, the existing methods rarely study the migration task between complex states. In light of the aforementioned issues, this paper puts forth AMDA motor fault diagnosis method based on multi-task alignment, with the aim of providing a solution to the aforementioned problems. The AMDA method employs a feature extractor comprising a multi-task one-dimensional convolutional layer, a batch normalization layer, and a pooling layer, to extract the higher-order features of the source and target domains. Subsequently, a combination of an adversarial-based approach and a distributional difference metric-based approach is utilized to reduce the distributional differences of data features. Finally, a multi-task learning approach based on gradient alignment is introduced to balance and optimize the three tasks: fault classifier, domain discriminator, and distributional difference metric. By reducing the conflicting gradients among the tasks, this approach ultimately enables the development of a domain adaptation fault diagnosis model for acoustic signals of electric motors based on multitask learning. Cross-operational state fault diagnosis tests are conducted under multiple test setups using the proposed AMDA method, and the test results demonstrate that the AMDA method effectively accomplishes the migration task between stable operational state (Stable), start operational state (Start), and European driving cycle state (NEDC) in the acoustic signal. Based on cross-operational state electric motor fault diagnosis tests, the highest diagnosis accuracies reach 91.49%. Furthermore, the performance of AMDA method is significantly higher than that of other methods in the two comparison tests, which offer valuable insights for research and engineering applications.

Abstract:Due to the absorption and scattering of light by water characteristics, underwater images usually present problems such as blurred details and low resolution. In order to improve the clarity of underwater images, an underwater image super-resolution reconstruction method that improves SRResNet is proposed. This method introduces the hybrid attention mechanism into the deep residual network to enhance the clarity of underwater images. Secondly, the structural similarity loss function is introduced to better protect image content, improve image quality, and make the training results more consistent with human visual perception. Experimental results show that the underwater image super-resolution reconstruction method based on improved SRResNet can effectively deal with problems such as blurred underwater images and low resolution. Compared to various other underwater image reconstruction methods on different datasets, this method improved the PSNR by 0.69 dB to 2.43 dB and the SSIM by 2.66% to 7.17%, demonstrating superior performance across all metrics.

Abstract:Aiming at the current problems of insufficient obstacle detection accuracy, slow detection speed, large number of model parameters and poor detection of small target obstacles in the complex environment of urban roads, an improved YOLOv8n lightweight urban driving road obstacle detection algorithm is proposed. Firstly, the MRObstacle urban road obstacle target detection dataset is produced to extend the types and numbers of obstacle detection; secondly, a new SPS_C2f backbone network is designed to improve the backbone network, to reduce the number of network parameters and to improve the detection speed, and the M_ECA attention module is added to the Neck portion of the network, to improve the network detection speed and the feature expression ability; thirdly, the BiFPN is integrated with a feature pyramid and a small target detection algorithm is added to the network. feature pyramid and adding a small target detection head to better capture the features of small-sized obstacles; finally, using the loss function MPDIoU that optimises the values of the bounding box width and height to improve the performance of the network bounding box regression. Compared with the original YOLOv8n algorithm, the mAP0.5 metric of this algorithm is improved by 2.04% to 97.12%, the FPS value is improved by 12.08 fps to 107.45 fps, and the volume of the network parameter is reduced by 10% to 2.73 MB.This algorithm improves the detection accuracy and speed while decreasing the number of parameters, and it can be better applied to the urban road obstacle detection task.

Abstract:Aiming at the current problems of limited stability and accuracy and high complexity of each solution in non-contact vital signs measurement, a non-contact vital signs measurement method based on matched filtering is designed to achieve low computational complexity and maintain optimality estimation. Five samples are tested in an office environment, and the results show the effectiveness of the proposed method to reduce the vital sign measurement errors due to people′s body movements in real environments. As an example, for sample 4, the variance of heart rate decreases from 2 825 to 82 in the smoothness design, and the root mean square error of heart rate decreases from 16 to 4 in the accuracy tracking calibration design. Clinical experiments are further compared with the current medical reference standards, and the results show that the respiratory rate error is within 1 bpm, while the heart rate measurements are better, which makes it potentially useful.

Abstract:Motor imagery (MI) EEG signals are more difficult to recognize due to the inclusion of long, continuous eigenvalues as well as their own strong individual variability and low signal-to-noise ratio. In this study, we propose a model that combines a convolutional neural network (CNN) with a Transformer, aiming to effectively decode and classify motor imagery EEG signals. The method takes the original multichannel motor imagery EEG signals as input, and learns the local features of the entire one-dimensional temporal and spatial convolutional layers by firstly performing a convolutional operation on the temporal domain of the signals in the first temporal convolutional layer, and then subsequently performing a convolutional operation on the null domain of the signals in the second spatial convolutional layer. Next, the temporal features are smoothed by averaging the pooling layers along the temporal dimension and passing all the feature channels at each time point to the attention mechanism to extract the global correlations in the local temporal features. Finally, a simple classifier module based on a fully connected layer is used to classify the EEG signals for prediction. Through experimental validation on the publicly available BCI competition dataset IV-2a and dataset IV-2b, the results show that the model can effectively classify MI EEG signals with average classification accuracies of up to 80.95% and 84.79%, which is an improvement of 6.45% and 4.31% in comparison to the EEGNet network, respectively, and effectively improves motor imagery evoked potential signals of the brain-computer interface performance.

Abstract:To address the challenges in mask detection for faces in dense crowd scenarios, particularly due to information loss from crowd occlusion, small detection targets, and low resolution, improved YOLOv8 algorithm for dense crowd mask detection is proposed. This approach replaces the FPN structure in YOLOv8 with a GD mechanism to solve the issue of missing cross-layer information transmission. The GD mechanism uses a unified module to collect and integrate information from all layers, enabling lossless cross-layer information transmission and enhancing the network′s feature extraction capabilities. The ODconv module is inserted into the GD mechanism to learn the information transmitted by GD along four dimensions, improving the model′s detection accuracy for low-resolution images and small targets. Additionally, a SCSBD is introduced to lighten the YOLOv8 detection head, which occupies a significant proportion, thereby balancing the model size. Experimental results show that the improved network has increased precision, recall, and mean average precision by 13.6%、1.5% and 6.9%, respectively, with an 81.1% accuracy in mask detection on faces. The model′s weight file is only 25 MB, and its size is between YOLOv8s and Gold-YOLO-S, achieving a balance between size and accuracy.

Abstract:To solve the problem of conveyor belt tear detection in the special operating environment of underground mines, a lightweight detection algorithm based on line laser assistance and improved YOLOv7 is proposed. Firstly, considering that the conveyor belt tear is mainly small targets, the largest detection layer is not needed, thus simplifying the network model to reduce the model size and the number of parameters. In addition, the dynamic nonmonotonic FM-based Wise-IoU loss function is adopted to make the model pay more attention to common quality samples and improve the model detection performance. Then, the LAMP pruning method is used to improve the model′s computing speed and reduce the computing complexity, achieving the lightweight of the detection network. The channel knowledge distillation is used to improve the model accuracy without loss, and finally, the model is accelerated by TensorRT to achieve faster detection speed. The experimental results show that compared with the benchmark model, the improved model has a parameter number and computing volume reduced by 86.8% and 49.2%, respectively, mAP@0.5:0.95 reached 62.4%, and the detection speed was improved by 151.0 fps, the model size was reduced from 71.3 MB to 12.8 MB. After the improvement, the model has improved the accuracy and real-time detection of conveyor belt tear faults.

Abstract:In the absence of disturbance information, the chattering effect in terminal sliding mode control of a permanent magnet synchronous motor (PMSM) becomes more pronounced. To reduce the impact of unknown uncertainties and disturbances on control performance, this paper presents an improved adaptive fast terminal sliding mode control method based on a disturbance observer. First, a mathematical model of the PMSM is developed, accounting for parameter uncertainties and load disturbances, and a nonsingular fast terminal sliding surface is designed to enhance the system′s response speed. A disturbance observer is then utilized to estimate system uncertainties and unknown disturbances, with an adaptive gain introduced in the sliding mode controller to compensate for estimation errors, achieving adaptive robust control without requiring a known upper bound for disturbances. This improved adaptive control strategy dynamically compensates for disturbances, strengthening the system′s adaptability to unknown disturbances. Simulation and experimental results demonstrate that, in the absence of disturbance information, the proposed method effectively suppresses chattering caused by sliding mode control, improves the robustness of the PMSM system, and significantly enhances control accuracy and dynamic performance.

Abstract:Aiming at the problem that traditional grey wolf algorithm is prone to local optimality in 3D path planning, an improved grey wolf algorithm is proposed in this paper. Firstly, the environment of the three-dimensional threat region is modeled, and the total cost function of UAV flight is specified under the constraint conditions. Secondly, chaotic sequences and quasi-reverse learning strategies were added to the initialization of grey wolf population, which increased the diversity of species and the search scope of unknown domain, and improved the adaptive weight factors to update individual positions, thus speeding up the convergence speed. Finally, in order to avoid falling into local optimization, particle swarm optimization algorithm is introduced to balance global development and local convergence. The experimental results show that compared with the other three typical path planning algorithms, the improved gray wolf algorithm can find a safe and feasible path, and has a stable optimization ability.

Abstract:Aiming at the problems of false detection and missing detection in the complex background of photovoltaic cell defect detection, an improved YOLOv8 based photovoltaic cell defect detection algorithm was proposed. Firstly, the bidirectional feature pyramid network is used as the feature fusion mechanism to achieve multi-scale feature fusion through top-down and top-down paths. Secondly, the context aggregation module is introduced into the neck network, and the context information of different receptive fields is obtained by using the cavity convolution of different cavity convolution rates, which helps the model to identify small targets more accurately, and thus improves the target detection performance of the model. Finally, the boundary frame loss function is optimized and its weight factor is adjusted continuously to improve the convergence speed and efficiency of the model. The experimental results show that compared with the detection network of YOLOv8 algorithm, the recall rate and average accuracy are respectively increased by 10.4% and 1.8%, and the detection frame rate reaches 270 fps, ensuring the lightweight requirements of real-time detection and subsequent deployment. The improved algorithm can carry out robust detection of photovoltaic cell defects under complex background.

Abstract:To address the problems of accuracy and efficiency limitations in traditional methods of studying vehicle trajectories and accelerate the promotion of digital road traffic management, this paper proposes a vehicle trajectory extraction method at intersections based on multi-target tracking optimization. First, based on the YOLOv8s algorithm framework, a multi-branch convolution strategy was introduced and an image processing method combining standard convolution and depthwise separable convolution was designed to improve the robustness of the model to different scenes and maintain a stable frame rate. Then, the loss function of the DeepSORT algorithm is improved by accurately quantifying the angle difference and distance loss to increase the convergence speed of the model and the accuracy of handling irregular objects. Finally, the accurate extraction of vehicle trajectories is ensured by deriving the conversion relationship between the pixel coordinate system and the real-world coordinate system. The experimental results show that the improved model has improved mAP, recall rate and MOTA by 2.9%、5.6% and 0.7% respectively compared with the original model, and the number of encoding transformations (IDS) has decreased by 64%. The frame rate can be kept stable during detection. And by deriving the conversion relationship between the pixel coordinate system and the real-world coordinate system, the vehicle′s trajectory information in the surveillance video can be accurately extracted. This provides methodological support for in-depth research on vehicle characteristics and road traffic risks, and has high practical application value.

Abstract:In order to solve the problems such as the large amount of YOLOv8 parameters affecting the detection speed, this paper proposes a lightweight leather defect detection algorithm based on the YOLOv8 framework by using automotive seat leather as a sample for defect detection on the surface of automotive seats. Firstly, the original backbone network of YOLOv8 is replaced with the lightweight network StarNet, which achieves the mapping of high-dimensional and nonlinear feature spaces through star arithmetic, thus demonstrating impressive performance and low latency with a compact network structure and low energy consumption. Secondly, the original detection head is replaced with a lightweight shared convolutional detection head, which allows for a significant reduction in the number of parameters through the use of shared convolution, making the model lighter so that it can be easily deployed on resource-constrained devices. Finally, the C2f module of the neck network is replaced by the C2f_Star module, which fuses feature maps of different scales while the network is more lightweight to improve the accuracy and robustness of target detection. Experimental validation of the model on the home-made HSV-Leather dataset shows that the improved YOLOv8-Leather detection model outperforms the YOLOv8n model. Compared to the YOLOv8n model, the improved model reduces the number of parameters by 57%, improves the detection speed by 20%, reduces the model weights by 52%, and reduces the computation by 53%. The experiment verifies the feasibility of the improved model in solving the problem of leather surface defect detection.

Abstract:Gait recognition is a key technology for lower limb exoskeleton robots, and accurate gait recognition plays a crucial role in the flexible control of these robots. To address the randomness in gait characteristics (such as walking speed and stride length) across different individuals and within the same individual, this paper proposes a gait phase recognition method based on an improved SECBAM-Densenet network model.Firstly, two inertial measurement units were placed on the tibia and the rectus femoris muscle of the thigh to collect gait data from 200 participants performing four gait tasks: walking forward, turning, ascending stairs and descending stairs. After filtering and resampling the data for preprocessing, the processed data were used as input to the proposed model. Finally, the SECBAM-Densenet model was used to classify the gait phases. The results show that the improved SECBAM-Densenet model achieved an average recognition accuracy of 95.76% in different gait phases within the same individual, which represents an improvement of 0.66% to 21.22% compared to other models. For different individuals, the recognition accuracy for each phase was higher than 94%.These experimental results indicate that the proposed model can be applied in the field of gait phase recognition, providing experimental reference for the flexible control of lower limb exoskeleton robots.

Abstract:In response to the problem of low accuracy in individual identification of communication radiation sources under channel noise interference, a communication radiation source individual identification method that integrates time-frequency characteristics is proposed by utilizing the difference in channel noise interference suppression effect of signal mapping to different time-frequency domains. Firstly, extract I/Q, power spectrum, and wavelet spectrum information from the radiation source signal, and fuse the time-frequency information of the signal through one-dimensional convolution in both horizontal and vertical directions; then, the channel attention module and spatial attention module are used to fuse time-frequency features; finally, M-ResNeXt network is used to achieve individual identification of radiation sources under channel noise interference. The experimental results show that under the interference of three channel noises, Gaussian white noise with a signal-to-noise ratio (SNR) of 15 dB, Rayleigh, and Rician, the recognition accuracy of the proposed time-frequency feature fusion method reaches 97.6%、97.7%、and 98.5% respectively. Even when facing unknown noise interference at an SNR of 15 dB, it can still achieve a recognition accuracy of over 97.7%. Therefore, the time-frequency feature fusion method can significantly improve the accuracy and robustness of individual communication radiation source identification.

Abstract:Aiming at the recognition problem caused by the complex background and the high similarity of commodity packaging in the vending machine scene, a commodity recognition method combining multi-scale attention mechanism and metric learning is proposed. Firstly, based on the ResNet hierarchical structure, multi-head self-attention is introduced to fully exploit the advantages of multi-scale feature extraction of convolutional neural network (CNN) and the global modeling ability of Transformer, and a new multi-scale hollow attention is designed to make the model focus on local features such as trademark shape and local texture in similar packaging, as well as context global features. Secondly, a down-sampling multi-scale feature fusion strategy is designed to effectively improve the multi-scale feature expression ability of the algorithm. Finally, ArcFace loss function is used to enhance the recognition ability of the model. In order to verify the effectiveness of the proposed method, a commodity data set in a real scene is constructed, which is collected by the top-view camera of the vending cabinet. The experimental results show that the MAP @ 1 accuracy of this method on the Commodity 553 dataset reaches 87.4%, which is better than the current mainstream recognition methods and can achieve more accurate commodity recognition.

Abstract:To address the low diagnostic accuracy of Alzheimer′s disease (AD) caused by the subtle complexity and spatial heterogeneity of brain lesions in structural magnetic resonance imaging (sMRI) of AD patients, a hybrid architecture that combines the strengths of convolutional neural networks (CNN) and Transformers is proposed for the AD diagnosis. First, a multi-view feature encoder is designed, in which a view local feature extractor with integrated hybrid attention mechanisms is employed to extract complementary information from the coronal, sagittal, and axial views of sMRI. The semantic representation of lesion regions is further enhanced through a multi-view information interaction learning strategy. Second, a cascaded multi-scale fusion subnetwork is designed to progressively fuse multi-scale feature map information, enhancing discriminative ability. Finally, a Transformer encoder is used to model the global feature representation of full-brain sMRI. Results on the Alzheimer′s disease neuroimaging initiative (ADNI) dataset show that the proposed method in this paper achieves classification accuracies of 94.05% for AD and 81.59% for mild cognitive impairment (MCI) conversion prediction, outperforming several existing methods.

Abstract:Aiming at the large model parameters and slow detection speed encountered by current lightweight target detection algorithms when applied to the task of detecting military aircraft in remote sensing images, this study proposes a lightweight detection algorithm for military aircraft targets based on YOLOv8n, named LeYOLO-MARs. The algorithm introduces an optimized inverted bottleneck module to replace the traditional bottleneck in the backbone network, reducing computational requirements while maintaining feature extraction capabilities and improving processing speed. In the neck network, a fast pyramid architecture is integrated to reduce the number of convolutional layers, enhance the efficiency of semantic information sharing, and decrease lock and wait times, while also considering limited parallelization opportunities and architectural complexity. A lightweight decoupled detection head, simplified through pointwise convolution, is employed, alongside the use of Inner-SIoU as the new localization regression loss function, which enhances the ability to learn from small target samples and accelerates the convergence of bounding box regression. Moreover, the algorithm incorporates a lightweight pyramid compression attention mechanism, effectively combining local and global attention to establish long-range channel dependencies. Experimental results demonstrate that the improved algorithm achieves a detection accuracy of 95.7%, 0.4% higher than the baseline model, while reducing model parameters by 43% and computational load by 63%, marking a notable improvement in detection performance compared to mainstream algorithms and enabling high-quality real-time detection of military aircraft targets.

Abstract:Underwater range-gated imaging technology is not affected by ambient light and has the advantage of long operating distances, making it a field of interest for many researchers. However, underwater gated images face issues such as uneven lighting distribution and high noise levels, which impair image clarity. In response to these challenges, this paper introduces an Enhanced Zero-DCE++ algorithm, building on the existing low-light enhancement algorithm Zero-DCE++. Initially, an improved kernel selection module is incorporated, replacing standard convolution and ReLU activation functions with depthwise separable convolution and ReLU6, to address overexposure issues in certain areas of underwater gated images. Furthermore, an improved HWAB half-wavelet attention module utilizing CBAM instead of the DAU dual attention unit is employed to differentiate between noise and real features in the wavelet domain, enhancing feature distinction and improving imaging clarity. Lastly, an ADNet noise reduction module is added to effectively suppress noise following low-light enhancement by Zero-DCE++. Experiments on a selfcollected underwater gated dataset demonstrate that the Enhanced Zero-DCE++ model achieves approximately 0.65 dB improvement in peak signal-to-noise ratio and a 0.23 increase in image information entropy compared to the Zero-DCE++ model, proving the model′s effectiveness and feasibility.

Abstract:To address the issue of inadequate target detection performance in the visual perception systems of autonomous vehicles, particularly under complex weather conditions such as fog and rain that introduce environmental noise, we propose a joint optimization target detection algorithm based on adaptive image denoising and multiple attention mechanisms(DMC-YOLO).An image denoising network has been constructed that combines the dark channel prior algorithm with ACE image enhancement technology to improve image quality in challenging weather conditions. Additionally, this network is integrated with the YOLOv8 backbone, utilizing SCDonw convolution to replace standard convolution. By incorporating point convolution and depth convolution, the aim is to reduce computational costs while obtaining richer down-sampling information.The SEAM attention module is employed to merge local and global information within the network. Furthermore, the SA detection head is introduced to emphasize contextual features, allowing for the retention of more detailed information. To enhance the network′s adaptability to various complex environments, linear interval mapping is incorporated into the loss function for reconstructing IoU.Experimental results indicate that, compared to the baseline model, the average accuracy of the improved algorithm increases by 2.9% while reducing the number of parameters by 15%. This effectively enhances the ability of autonomous vehicles to recognize targets in complex environments.The deployment outcomes on EC-R3588SPC and Nvidia Jetson NX edge devices are promising, fulfilling real-time detection requirements even under challenging weather conditions.

Abstract:Aiming at texture problems, contour similarity among fresco image characters, large differences in fresco character features in different scenes, complex background noise, and confusing classification, an improvement strategy for ResNet convolutional neural network is proposed. Firstly, the larger 7×7 convolutional kernel in the input layer of the model is separated into three series-connected 3×3 small convolutional kernels stacked in the backbone, and 2×2 average pooling and maximum pooling are used for add feature fusion to replace the original maximum pooling operation, which enhances the model′s representative ability. Secondly, a multi-scale efficient spatial channel attention module is designed, based on the ECA channel attention module, the spatial attention module is connected in series, and the original 3×3 convolutional kernel in the spatial module is replaced by the SK attention module, which fuses the multi-scale information to capture the global long-distance dependency, and reduces the interference of background noise. Finally, a cellular aggregation structure is proposed to perform ADD operation on the output information in the neighboring block blocks as inputs to the subsequent layers, capturing both low-level and high-level features to enhance the circulation of contextual information. The experimental results show that the model achieves 96.51%、96.65%、96.67% and 96.63% in accuracy、precision、recall and F1 value, respectively. Relative to the original model ResNet-18 accuracy is improved by 9.76%, and compared with mainstream classification algorithms classification accuracy, generalization ability, and stability are all improved, which can efficiently and accurately identify the type of mural belonging to the mural, which is of significant value for cultural heritage preservation and art history aspects of the research.