Abstract:To study the variation characteristics of sheet resistance of thin film samples under variable temperature conditions, a device integrating a Bayesian optimization fuzzy PID temperature control system and a four-probe method was designed. The Bayesian optimization algorithm was used to find the optimal parameter combination for the fuzzy controller, significantly improving the temperature control accuracy of the system and greatly reducing the experimental cost of the parameter optimization process. A highly stable constant current source circuit and a high-precision digital voltmeter were designed for the four-probe system to ensure stable current supply and accurate voltage measurement even when the resistance of the thin film changes by two to three orders of magnitude. Through an efficient two-way communication protocol between the embedded system and the host computer, the duty cycle of PWM, temperature and voltage data were synchronously transmitted in real time, achieving high-precision temperature control and real-time calculation of the sheet resistance of the thin film samples.

Abstract:Vehicle geomagnetic navigation faces a key issue of magnetic interference affecting navigation accuracy during dynamic driving. Traditional fixed-parameter compensation methods struggle to adapt to interference changes under different motion states. This paper reveals the limitations of a single fixed-parameter compensation model by analyzing the relationship between motion state and magnetic compensation effect, providing a theoretical basis for the development of adaptive compensation methods. The CNN-SRU motion state recognition model constructed in the study achieved a recognition accuracy of 99.61%, with training efficiency improved by 12.8%~28.4% compared to the comparative model, and inference delay reduced by 25.4%~38.5%. Based on the recognition results, the compensation performance of the single ellipsoid fitting model was systematically evaluated, and significant differences in compensation effects under different motion states were found: The standard deviation after compensation was 49.39 nT for uniform linear motion, which exhibited the best performance; the standard deviation after compensation was 533.35 nT for steering motion due to complex interference; and the standard deviation after compensation reached 147.98 nT for acceleration motion due to strong transient interference. The research indicates that motion state is a key factor affecting the magnetic compensation effect, and fixed-parameter models cannot meet the requirements of all operating conditions. The “state recognition-compensation evaluation” framework established in this paper provides theoretical support and technical paths for the development of adaptive magnetic compensation methods.

Abstract:Steel surface defect detection plays a vital role in industrial quality control, yet complex textures and multi-scale characteristics increase the difficulty of accurate detection. To address this, a lightweight detection network named MEA-YOLO, based on multi-scale edge enhancement and attention fusion, is proposed. The approach replaces the YOLOv11 backbone with StarNet to reduce computational complexity, introduces a multi-scale edge information enhancement module combined with a dual-domain selection mechanism to strengthen boundary and contextual features, and incorporates a spatially enhanced feedforward network in the detection head to improve fine-grained recognition. Experimental results show that the proposed method achieves an mAP50 of 74.58% on steel defect detection tasks, outperforming YOLOv11n by 2.01%, while reducing the number of parameters and GFLOPs by 14.3% and 10.1%, respectively. Additional evaluations on the GC10-DET dataset further demonstrate consistent improvements in both accuracy and inference speed, confirming the model′s robustness, generalization capability, and suitability for real-time industrial inspection.

Abstract:In response to the demand for high-bandwidth short-range signal transmission of multi-physical parameters in turbomachinery under conditions of high rotational speed, significant axial movement, and spatial constraints, this paper proposes a design approach using a 900 MHz/2.4 GHz dual-band circumferentially distributed transmitting antenna array and a ring-shaped receiving antenna. Independent and coupled solution models for the receiving antenna and transmitting antenna array were established. A transmission reliability evaluation method for rotational and varying-distance conditions was proposed. By analyzing the impact of microstrip line width and dielectric constant on antenna performance, the method thereby confines frequency deviations within the operational band. Simulation and experimental results show strong agreement: The 900 MHz transmitting antenna achieves an impedance bandwidth of 10 MHz, the 2.4 GHz antenna provides an impedance bandwidth of not less than 90 MHz, and the receiving antenna′s impedance bandwidth fully covers the frequency range from 900 MHz to 2.4 GHz. When the relative rotational speed does not exceed 11 000 rpm and the spacing remains within 7±4 mm, stable and reliable transmission of turbomachinery parameters can be achieved.

Abstract:In response to the problems of insufficient practical orientation and single evaluation dimensions in traditional equipment testing capability evaluation methods, a comprehensive evaluation system was constructed with combat effectiveness as the core orientation, which includes three dimensions: Experimental design, implementation effectiveness, and practical assessment. On the basis of this evaluation system, satisfaction is used as a standardized evaluation criterion to solve the problem of normalizing multi-source heterogeneous data. The analytic hierarchy process is applied to weight the contribution of combat effectiveness, and the system contribution rate is introduced to iteratively revise the evaluation model. A equipment testing capability evaluation model with combat mapping characteristics is established. Through this method to evaluate the contribution of a test combat power generation, it can be concluded that the contribution of the test capability to combat power generation is 71.5, and it is relatively satisfied with the current test capability, which is consistent with the actual situation. After the test improvement, the contribution to combat power generation can be increased by 18.88%, which effectively solves the problem of the value correlation between the test capability and the contribution to combat power generation, and provides theoretical support and practical methods for the construction of new equipment test evaluation system and the transformation of combat power generation mode.

Abstract:The intensive care unit serves as a central hub for treating critically ill patients, with the essential responsibility of providing continuous monitoring and care for severe cases. However, the current vital sign monitoring systems used in most domestic ICUs mainly depend on wired equipment, which comes with various limitations. This paper presents a wireless vital sign monitoring system that can be comfortably worn on the patient′s wrist and fingertips to address this issue. This system is designed to collect real-time data on the patient′s body temperature, heart rate, and blood oxygen saturation levels. The collected data is then uploaded to both a mobile phone and a cloud-based database for easy access and display. Compared to the existing wired monitors, this device is compact, easy to wear, and enhances patient comfort and mobility. It also reduces physical contact with the equipment, decreasing the infection risk. During testing, one individual wore the wireless monitoring device alongside a Huawei smart band and a medical-grade finger-clip heart rate and blood oxygen monitor (considered the standard) while resting. The results indicated that the standard error for heart rate data from the wireless monitoring device was 0.15, while for blood oxygen data, it was 0.065. In contrast, the Huawei smart band showed a standard error of 0.48 for heart rate data and 0.13 for blood oxygen data. These findings suggest that the wireless monitoring device provides greater measurement accuracy than the Huawei smart band and outperforms the finger-clip heart rate and blood oxygen monitor in terms of continuous monitoring and data recording capabilities. This advancement is significant for the evolution of intelligent ICUs.

Abstract:To address the performance bottleneck of contour-based gait recognition, where over-reliance on global representations leads to sharp degradation under strong appearance interferences like clothing changes, this paper proposes a model integrating structure awareness and dynamic attention. The model aims to elevate the recognition paradigm from matching variable contours to understanding intrinsic motion patterns. To achieve this, this study construct a dual-path parallel framework: First, a structure-aware path precisely models key local body regions; second, a frequency-decoupled dynamic attention mechanism is introduced to adaptively enhance the most discriminative feature channels against gait phase variations; finally, a deep semantic fusion module synergizes local structural information with global representations at multiple scales to generate a final feature with both stability and discriminative power. Experimental results show that the model achieves an average accuracy of 89.9% on the CASIA-B dataset, with 11.0% improvement over the baseline under the changing-clothes condition, and a Rank-1 accuracy of 89.5% on the large-scale OU-MVLP dataset. This study confirms that by synergizing local structure perception and global feature enhancement, the proposed model effectively improves the robustness and accuracy of gait recognition under complex appearance interferences.

Abstract:Currently, traditional Braille teaching tools such as Braille books can hardly meet the digital reading needs of the blind or visually impaired (BVI), and the developed refreshable Braille displays (RBDs) still need to be further optimized in terms of volume, locking force, energy consumption and other aspects. To help BVI read digitally conveniently, a low-power Braille display module is proposed, and a wearable rotating RBD is designed by combining a rotating structure. Firstly, an electromagnetic actuator is designed based on the principle of electromagnetic drive. This actuator can rely on electromagnetic characteristics to lock the raised Braille dots without energy consumption. Then, through the force analysis and finite element analysis of the Braille dot actuator in different states, the key parameters of the Braille dot actuator that meet the design requirements are determined. The performance test results of the device show that the average refresh frequency of the Braille dot actuator is 19.3 Hz, and the average locking force is 166 mN, which can meet the needs of BVI for touch reading. User experiments show that the average accuracy of blind subjects in recognizing Braille with the wearable rotating RBD reaches 94.66%. Therefore, the wearable rotating RBD provides a convenient digital Braille reading method for BVI.

Abstract:Refining units in southern coastal China operate in complex service environments where multiple environmental factors are strongly coupled, making metallic equipment prone to accelerated corrosion during long-term operation. High-accuracy prediction models are therefore needed to support equipment health management. To address the limited availability of corrosion data in refining units (target-domain sample size N=12), this study proposes an atmospheric corrosion-rate prediction method termed TL-BiLSTM-Attention, which integrates transfer learning with a BiLSTM-Attention network. First, a source-domain dataset is constructed from the public MICAT corrosion dataset. Then, orthogonal experiments are conducted to optimize the hyperparameters of the BiLSTM-Attention network to improve the performance of the pre-trained model. Finally, the optimal pre-trained model is transferred to the local corrosion dataset of the refining units, and fine-tuning is performed under different layer-freezing strategies. Results show that, in the source domain, the BiLSTM-Attention model achieves stronger fitting capability and lower errors than LSTM and BiLSTM. In the target domain, TL-BiLSTM-Attention improves the prediction accuracy of atmospheric corrosion rates for materials such as Q235, 304, 316L, 5A06 and different metals require differentiated freezing strategies to obtain the best performance. This study verifies the predictive effectiveness and engineering applicability of TL-BiLSTM-Attention under few-shot conditions and provides a data-driven tool for corrosion assessment in refining units.

Abstract:Addressing the challenges faced by autonomous driving systems in nighttime scenarios, such as low background contrast, image blurring, and complex lighting interference, which lead to poor detection performance, this paper proposes a lightweight RDSM-YOLO model based on the YOLOv8 network. The model focuses on comprehensively enhancing key feature extraction and fusion in low-light scenarios. First, RFAConv is introduced into the backbone and neck networks, utilizing a receptive field attention mechanism to adaptively highlight key spatial features; second, the DynamicConv module is used to reconstruct the C2f module, enabling dynamic aggregation of convolutional kernels to enhance feature expression without increasing FLOPs; simultaneously, the lightweight SPPELAN module replaces the traditional SPPF to fuse multi-scale contextual information; finally, the loss function is upgraded from CIoU to EIoU, explicitly decoupling bounding box geometric elements to accelerate convergence and improve localization accuracy. Experimental results show that RDSM-YOLO achieves a mAP50 of 70% for nighttime vehicle detection on the BBD100k dataset, improving by 1.4% over YOLOv8 while maintaining a model parameter count of only 3.04 M. This paper demonstrates that the proposed model achieves both lightweight design and high accuracy in nighttime vehicle detection, providing a reference for improving nighttime autonomous driving performance.

Abstract:Targeting the mobility assistance needs of people with walking difficulties in environments dense with obstacles like museums and exhibition halls, this study proposes an autonomous navigation path planning method. The method combines a global A* algorithm with a local dynamic window approach DWA. First, an improved A* algorithm designs an optimized heuristic function and a path smoothing strategy. This improved algorithm generates the global optimal path. Second, the study employs an adaptive weighting DWA algorithm to handle dynamic environments. This DWA variant adds a global path evaluation function and a braking distance evaluation function to the original DWA. These additions significantly enhance safety and motion continuity within complex, narrow passages. To verify effectiveness, the study constructs a test platform simulating a museum environment in a laboratory setting. Evaluation utilizes metrics including trajectory smoothness, collision rate and number of traversed nodes. Experimental results demonstrate that compared to traditional A* and original DWA algorithms, this method achieves global path optimality. It also increases the dynamic obstacle avoidance success rate. Furthermore, the method reduces trajectory jitter amplitude. This approach provides reliable technical support for the safe movement of special groups in constrained indoor environments.

Abstract:To address the decline in detection accuracy caused by complex water surface backgrounds such as wave interference, light reflection, and object occlusion, an improved YOLOv11n-based water surface garbage detection algorithm is proposed. First, the Bottleneck structure in C3k2 is replaced with the FCM module to construct the C3k2-FCM module, mitigating the loss of spatial position information during downsampling. Second, a FPSC-SCSA module is designed by introducing shared convolutional layers with different dilation rates and the SCSA mechanism to replace the SPPF module, enhancing the model′s focus on key regions and reducing the loss of crucial information. Third, a BIFPN-V2S module is developed by replacing the neck aggregation network with a multi-scale feature fusion network embedded with the SDI module from U-Net V2, expanding the receptive field and strengthening global contextual interactions. Finally, an SGSV-IoU loss function is designed to improve boundary shape and detail representation, enhancing the localization of irregular floating objects. Experimental results on a self-built water surface garbage dataset show that, compared with the original YOLOv11n model, the improved algorithm increases mAP50 by 2.8%, while reducing parameters and computation by 0.81 M and 0.1 GFLOPs, respectively, demonstrating the effectiveness of the proposed method.

Abstract:To address the technical bottlenecks of poor detection accuracy and robustness in small object detection caused by insufficient feature information and low-resolution feature maps, this paper proposes a small object detection method based on high-resolution feature-guided learning. The method adopts an improved YOLOv11 algorithm structure based on context aggregation pinwheel convolution and constructs a dual-channel detection framework consisting of a high-resolution detection branch and a low-resolution detection branch. During the training process, the high-resolution detection network guides the learning of the low-resolution detection network, alleviating the problem of insufficient semantic information of small objects in low-resolution images. A multi-scale feature alignment loss function with weighted multi-loss functions is introduced in the middle layer of the dual-channel network to enhance the expressive ability of small object features. Experimental results show that the proposed method achieves a 4.11% improvement in mAP50 and a 4.07% improvement in mAP50:95 compared to the original YOLOv11 on the PASCAL VOC 2012 small object dataset; on the Visdrone2019 dataset, the mAP50 increases by 2.24% and the mAP50:95 increases by 1.50% compared to the original YOLOv11.

Abstract:Non-intrusive load monitoring is a technology that identifies operating electrical appliances and their energy consumption by analyzing voltage and current variations on the main power bus. With the continuous increase in the types and quantities of electrical loads, extracting unique load features and establishing efficient non-intrusive load monitoring classification models have become particularly important. This paper proposes an image feature enhancement method centered on multi-voltage-current trajectory fusion and background feature mapping, and applies it to the ResNet18 network via a transfer learning strategy. By means of multi-V-I trajectory fusion and image background mapping, the accuracy of load identification is improved, thus achieving efficient classification of non-intrusive loads. Different from traditional methods, this paper for the first time proposes a differential fusion strategy of full-wave and filtered trajectories, which enhances the uniqueness of load features. Additionally, by mapping multiple steady-state features in the image background, the representational capability of the images is further improved. Experimental results demonstrate that the load identification accuracies of this method on the PLAID2014, PLAID2017 and PLAID2018 datasets are increased to 98.67%, 97.53% and 98.64% respectively, exhibiting significant advantages over existing models.

Abstract:Deep vein thrombosis can potentially give rise to severe complications like pulmonary embolism, which poses a threat to the life safety of patients. Therefore, early prediction of DVT risk holds significant clinical implications. However, current DVT risk prediction methods mainly focus on only predict using either single-text or single-image data, and there are few studies which integrate these two types of modal data for DVT risk prediction. To address these challenges, this study combines the Mamba state space model with multimodal fusion and proposes a novel DVT risk prediction method based on Mamba self-attention and multimodal fusion for the first time. This method takes the patient′s ultrasound images and structured text data such as medical history and laboratory test indicators as multimodal input data. Firstly, a dual-channel feature encoding framework is constructed, which uses ViT to capture the features of ultrasound images and DNN to obtain the features of structured clinical data. Then, this paper proposes a multimodal feature fusion framework based on Mamba self-attention. This framework first concatenates the image and text features to obtain the joint features, and then uses the original Mamba to train the joint features to obtain the multimodal fusion features. Subsequently, Mamba self-attention, feedforward network, and CNN are designed to extract and fuse global and local, high-level and low-level features of multimodal data, thereby preserving the original multimodal features from multiple perspectives. Finally, multi-level MLP is used for feature dimension reduction to obtain the DVT prediction results. Comparative experiments were conducted on a clinical dataset with 13 other combined models. The results show that this model outperforms the others, with an AUC of 0.912, an average improvement of 11.97% compared to the single structured data model, and an average improvement of 13% in F1 score. Compared with the traditional single image data model, the AUC is improved by an average of 14.7%, and the accuracy and F1 score are both increased by more than 20%. Among the multimodal comparison models, this model outperforms the ResNet and Transformer fusion model (AUC=0.871) in terms of accuracy, precision, recall, and F1 score by approximately 6%. Compared with the same-structured Transformer hybrid model, the AUC and the other four performance evaluation indicators as well as the model inference speed are all improved by more than 20%. The results indicate that the model proposed in this study provides strong support for the early prevention and prediction of DVT and has good application prospects and clinical value.

Abstract:Based on the modular multilevel converter (MMC), flexible direct current (DC) transmission technology has been widely adopted in China′s power system. In recent years, from the Yu-Hu Project to the Zhangbei Project, high-frequency oscillation phenomena have occurred in multiple voltage-source converter based flexible DC transmission projects both domestically and internationally, jeopardizing the safe and stable operation of the system. To address this issue, this paper proposes a high-frequency oscillation detection method for flexible DC transmission systems based on wavelet packet decomposition (WPD) and unscented Kalman filter (UKF) tracking. First, the filter bank structure of fast wavelet packet decomposition (WPD) is employed to accurately partition the grid-connected current signal of the flexible DC system into frequency bands. Then, a threshold is set based on the wavelet packet coefficients of each node at every layer. If the threshold is exceeded, it is determined that a high-frequency oscillation may be present. The wavelet packet coefficients that exceed the threshold are reconstructed to obtain a time-domain signal containing only specific frequency components. Subsequently, the unscented Kalman filter (UKF) is applied to track this signal, yielding dynamic frequency tracking curves and dynamic amplitude tracking curves. Finally, an MMC-based flexible DC transmission system model was built on a simulation experimental platform to generate high-frequency oscillations. The results verify the sensitivity and accuracy of the proposed method in detecting high-frequency oscillations, meeting the detection requirements for subsequent oscillation suppression.

Abstract:The backward projection algorithm belongs to the time-domain algorithm in synthetic aperture radar imaging algorithms and is often used as a comparative algorithm for mixed domain or frequency-domain algorithms. This article designs a one-dimensional sparse topological linear array equivalent to a one-dimensional dense linear array through phase compensation, and derives the matched filtering algorithm and wavenumber-domain backward projection algorithm. The numerical simulation results show that the simulated multi-point target point spread function has a horizontal and vertical direction of about 3.5 mm and a distance direction of about 12.6 mm, verifying the correctness of the matched filtering algorithm and the wavenumber-domain backward projection algorithm. By building a physical system, matching filtering algorithm, wavenumber-domain backward projection algorithm, and backward projection algorithm were used to image the line pair resolution plate, with horizontal and vertical resolutions of 4 and 5 mm, respectively. The simulation results were in good agreement with the theoretical resolution. Finally, for the simulated imaging of the actual target, the matched filtering algorithm took 4.461 s, the wavenumber-domain backward projection algorithm took 54.477 s, and the backward projection algorithm took 684.985 s, which is consistent with the theoretical complexity of the three algorithms. The median filtering gradient similarity value in the images of the three algorithms is about 0.9, and the image contrast value is about 6.6, verifying the effectiveness of the matched filtering algorithm and the wavenumber-domain backward projection algorithm.

Abstract:To address the limitations of existing image splicing and tampering detection methods, such as restricted receptive fields, single-scale feature extraction, and limited extraction capabilities, this paper proposes a multi-scale gated fusion algorithm for image tampering and splicing detection. First, a dual-encoder, single-decoder architecture network is designed: the two encoders utilize standard convolution and dilated convolution to capture features at different scales, while the decoder employs standard convolution. Second, at the shallow layers of the network, the features extracted by the two encoders are added element-wise to fuse dual-path information, which is then passed to the decoder via skip connections to enhance feature representation capabilities. Finally, at the end of the encoder, a multi-scale adaptive gated fusion module is employed to adaptively fuse the local and global features captured by the dual encoders, thereby reducing redundant information and highlighting important features. Experimental results show that the proposed method achieves F1 score improvements of 9.62%, 3.29%, 4.75% and 2.5% on the three public datasets CASIA1.0, CASIA2.0, IMD2020 and a self-created synthetic dataset, respectively; in comparative experiments, the proposed method outperforms other methods in overall detection results; in robustness experiments, the proposed method demonstrates high accuracy and robustness in handling complex scenes and diverse data, effectively detecting tampered regions with performance superior to other methods. The above results indicate that the proposed method provides a robust technical foundation and new research directions for studies and applications in the field of image security.

Abstract:To address the limited robustness of traditional Siamese networks in complex scenarios and the high computational demands of Transformer-based architectures, this paper proposes a novel object tracking algorithm based on supervised feedback and a lightweight Transformer. First, a supervised feedback module is designed to incorporate task-relevant feedback during feature extraction, guiding the network to focus more effectively on target regions, thereby enhancing feature discriminability and suppressing background interference. Second, a lightweight Transformer structure is constructed, which maintains strong global modeling capabilities while significantly reducing computational complexity and parameter overhead, achieving a favorable balance between performance and efficiency. Finally, an adaptive template update mechanism is introduced to dynamically adjust the template content based on the current frame′s object state and scene variations, improving adaptability to target appearance changes and mitigating tracking drift. Experimental results on multiple mainstream public datasets show that the proposed method outperforms existing advanced algorithms in terms of both robustness and real-time performance.

Abstract:To address the issues of low edge location accuracy, insufficient multi-scale feature extraction capability, and complex background interference in the detection of small target defects in transmission lines, a target detection algorithm based on STDD-YOLO is proposed. This algorithm enhances the high-frequency feature extraction capability by designing an edge-spatial feature enhancement module, improving the positioning accuracy of defect boundaries and enhancing the perception of high-frequency information such as edges and contours. It replaces the standard convolution in the original Bottleneck with efficient multi-kernel convolution to enhance the network′s detection performance for multi-scale small targets and solve the limitations of the C3k2 structure in utilizing detail information. A shared conv group norm head is designed to suppress background noise interference, enhance the feature expression ability of small targets, effectively improve the robustness of model detection in complex environments, and avoid false detection and missed detection caused by the low saliency of defect targets in complex backgrounds. Experiments show that this algorithm can significantly improve detection accuracy.

Abstract:To address the challenges of low detection accuracy for small-target vehicles/pedestrians and high computational complexity in autonomous driving scenarios, this paper proposes DSWR-YOLO—an improved lightweight object detection algorithm based on YOLOv10n. First, a DWR module is introduced, where its dilated residual structure is enhanced to DSConv. The SimAM attention mechanism is embedded to reconstruct the C2f module, strengthening feature retention for small targets. Second, an additional 160×160 detection layer is incorporated. A novel dynamic detection head, Detect-dyHead-P2, is designed using multi-expert fusion and separation combined with DynamicConv, significantly reducing model parameters while improving small-target detection capability. Finally, the Focaler-SDIoU loss function is integrated to dynamically adjust loss weights, mitigating sample imbalance and unstable bounding box regression.Validation on the VisDrone2019 dataset demonstrates that DSWR-YOLO achieves a 25.9% reduction in parameters and 33.3% decrease in FLOPs, while improving mAP@0.5 by 3.7%, Precision by 2.9%, and Recall by 3.3%. This delivers enhanced accuracy with reduced computational costs, making it suitable for resourceconstrained embedded devices. Generalization experiments on the UA-DETRAC dataset show that DSWR-YOLO outperforms the baseline by 0.9% in mAP@0.5, 1.2% in Precision, and 2.0% in Recall, confirming its robust generalization capability.

Abstract:Aiming at the problems of low efficiency and insufficient positioning accuracy in traditional sub-pixel edge detection methods, this paper proposes a sub-pixel edge detection algorithm based on Prewitt operator and Franklin moment fusion. Firstly, the Prewitt operator is used to realize the pixel-level initial positioning of the edge, and then the sub-pixel edge analysis model based on Franklin moment is constructed. The multi-order moment convolution is used to extract the edge features. Combined with the orthogonality and rotation invariance of the moment function, the relationship between the moment values before and after the image rotation is established, so as to accurately calculate the sub-pixel coordinates of the edge points, and the effective screening of the edge points is completed by the improved discriminant conditions. Finally, the least square method is used to fit the edge. The experimental results show that the average error of the jet port detection is 0.002 6 mm, the average detection time is 0.805 s, and the positioning accuracy is higher, the noise resistance is stronger, and the universality is better. It can better solve the problems of low positioning accuracy and slow efficiency of image edge detection.

Abstract:To address the challenges prevalent in polyp segmentation, such as co-occurrence phenomena, boundary ambiguity, and under-segmentation, a novel multi-scale attention fusion network, MSBRAU-Net++, is proposed. It adopts multi-scale gated attention fusion to create an interactive structure between multi-scale features and attention, processing contextual information to enhance foreground feature responses and suppress background interference, thereby significantly improving the ability to distinguish polyps from similar tissues. By utilizing a hybrid spatial channel module, it addresses the issue of boundary ambiguity through deep feature calibration and local detail recovery, thereby enhancing the precision of edge segmentation. A novel multipath feature aggregation block is designed to fuse low-level details with high-level semantic features, preventing information loss and ensuring the integrity of the segmentation results. MSBRAU-Net++ was evaluated on the Kvasir-SEG and CVC-ClinicDB datasets, achieving IoU scores of 84.65% and 88.87%, and DSC scores of 90.63% and 91.99%, respectively. The experimental results demonstrate that MSBRAU-Net++ outperforms other comparative models and can accurately segment images, showing particularly significant results in segmenting regions with complex boundaries and small polyps.

Abstract:To address the challenges of small object size, severe occlusion, and complex backgrounds in UAV aerial images, traditional object detection algorithms often suffer from missed and false detections. To improve detection accuracy, this paper proposes a small object detection algorithm, YOLOv11n-AFD, based on YOLOv11n by integrating attention and feature modulation mechanisms. The method incorporates a Spatial Strip Attention (SSA) module, a Modulation Fusion Module (MFM), and a Manhattan Feature Enhancement (MFE) module to comprehensively enhance the model′s perception and semantic representation of small objects. Within the unchanged YOLOv11n framework, the SSA models horizontal and vertical spatial dependencies to strengthen structural awareness; the MFM refines feature fusion through channel modulation to highlight key information; and the MFE reinforces geometric features while suppressing background interference for deeper enhancement. Experimental results show that YOLOv11n-AFD achieves improvements of 1.8% in precision, 0.8% in recall, and 1.4% in mAP@0.5 over the original YOLOv11n, with mAP@0.5:0.95 increasing to 21.5%, demonstrating superior performance compared with other algorithms.

Abstract:Dynamic point clouds have significant value in cutting-edge applications such as immersive communication and autonomous driving, and their efficient compression is key to achieving real-time transmission and storage. Although rule-based and learning-based approaches have made progress in point cloud geometry compression, existing methods are still insufficient in leveraging inter-frame correlations in dynamic sequences. This paper proposes a dynamic point cloud geometry compression method based on feature-domain nearest neighbor search and concatenation, extending the multi-scale sparse representation framework to dynamic scenes, and introducing multi-scale temporal priors to enhance inter-frame conditional coding. Specifically, by extracting hierarchical features from the reconstructed reference frames and performing nearest-neighbor search and concatenation with the current frame features in the feature domain, spatiotemporal contextual information across spaces is constructed, thereby enabling more accurate estimation of voxel occupancy probabilities. This method transmits only partial features at the encoding end, and the decoding end reconstructs the temporal priors using reference frame information, significantly improving compression efficiency. The experiment was conducted on a standard dataset following the MPEG general test conditions. The results indicate that the method proposed in this paper achieves significant BD-Rate gains of over 10% in terms of D1-PSNR and D2-PSNR on multiple test sequences compared to existing rule-based and learning-based compression methods, particularly demonstrating superior rate-distortion performance across a wide range of bitrates. The test results validate the effectiveness and advancement of the algorithm proposed in this paper for dynamic point cloud geometry compression.