Abstract:The existing deep learning methods don’t make full use of the prior information such as semantic and texture information in intact area of mural restoration， which results in poor restoration results， so a mural restoration algorithm based on semantic prior and texture enhancement guidance is proposed. Firstly， a semantic prior learning module is designed， which maps the original mural semantic features to a semantic prior learner through pixel folding operations. The original semantic features are used to guide the repair of incomplete features， gradually reducing the difference between damaged and original semantic features. Then， a texture enhancement module is designed， which enhances texture details by fusing contextual information modules and fusing them to complete the restoration of mural texture features. Finally， an aggregate bootstrap module is designed， which integrates the semantic prior repair and texture enhancement results， decodes them to the original resolution， and completes the repair of damaged murals through adversarial games with Markov discriminators. The digital classification restoration experiment of Dunhuang murals shows that the proposed method outperforms the comparative algorithm in both subjective and objective evaluations， achieving better restoration results.

Abstract:In response to several key challenges faced by the existing tone mapping techniques in practical applications， such as insufficient stability of mapping results， difficulty in balancing the natural aesthetics of images， and limited adaptability to complex lighting environments and diverse scene types， this paper proposes a tone mapping method based on multimodal learning. The goal is to acquire cross-modal supervisory information through the shared semantic space of text and images， aiming to achieve more accurate， natural， and universally applicable tone mapping. By leveraging the text-image matching information from large text-image models to assist in unsupervised training， the method effectively suppresses the occurrence of underexposed and overexposed areas， avoiding the training instability and complexity issues present in generative adversarial methods and contrastive learning. Experiments demonstrate that the proposed tone mapping method displays superior performance across multiple open benchmark datasets. Compared with the existing mainstream tone mapping algorithms， this method not only maintains the overall lighting atmosphere of images but also more effectively suppresses overexposed areas， enhances underexposed areas， retains rich color details， and enhances visual hierarchy， with stronger adaptability to various lighting conditions and scene types. Moreover， this work also confirms the significant potential of multimodal learning in foundational vision tasks.

Abstract:To address the limitations of current digital halftoning algorithms， such as slow processing speed and suboptimal halftoning effects， a data-driven halftoning framework is proposed. By introducing the Gumbel-Softmax reparameterization strategy， the non-differentiability issue caused by discrete halftone selection is resolved， enabling unbiased gradient estimation during network backpropagation. To further enhance the halftoning effects， a novel blue noise loss function is designed to optimize the distribution of halftone dots. Additionally， a Patch-wise Confidence Aggregation module is introduced to incorporate spatial correlations between pixels， allowing the network to focus more on pixel interactions during training. Based on these strategies， a label-free， self-supervised， differentiable halftoning framework is constructed by optimizing the expected value of the halftone quality metric. Experimental results demonstrate that the proposed method， without requiring image labels， can generate high-quality halftone images and maintain high processing speed and low parameter complexity， effectively preserving local structural information and texture details. Moreover， this framework can be flexibly extended to multi-level halftoning to accommodate the requirements of multi-level printheads.

Abstract:Deep learning-based methods， particularly those employing Siamese network structure， are widely used in remote sensing image change detection. However， extensive experiments reveal that these methods often suffer significant performance degradation when the order of input images is altered. Notably， the ChangeFormer method exhibits a 79.86% drop in the intersection over the union （IoU） metric on the LEVIR-CD dataset， indicating a lack of chronological robustness that severely impacts the practical utility of the change detection model. To address this issue， this article proposes a novel change detection method called chronologic invariant network （CINet）， which integrates chronological alignment module and cross-layer feature mixing module. During the feature extraction phase， a chronological alignment module is introduced. This module employs spatially cross-mix feature maps and chronologic reconstruction to reduce temporal information discrepancies between the two branches at the feature level. Additionally， a cross-layer feature mixing module is designed to blend deep and shallow features using a full-scale connection and difference-guided approach， effectively utilizing feature information from every layer of both branches to improve change detection accuracy under varying input sequences. Experimental results on the LEVIR-CD dataset show that CINet achieves a recall of 90.63% and an IoU of 84.13%， which represent improvements of 1.83 percentage points and 1.65 percentage points over ChangeFormer， respectively. Results from multiple datasets further demonstrate that the proposed method consistently maintains high change detection performance and robust chronological invariance， even when the order of input images is altered， outperforming other methods.

Abstract:Underwater images from different scenes often exhibit complex and non-uniform degradation due to factors such as light absorption by water and the scattering effects of suspended particles. Even within the same image， the degradation degree varies across regions due to differences in scene depth. The most existing underwater image enhancement methods fail to specifically address the non-uniform degradation， leading to poor enhancement results. To solve this issue， this paper proposes an iterative underwater image enhancement network （IUIENet） based on degration distribution perception. IUIENet consists of three modules： a pre-enhancement module， a degradation distribution estimation module and an image refinement module. The pre-enhancement module initially estimates the enhancement result， while the degradation distribution estimation module and the image refinement module optimize the enhancement results using iterative cooperation. Experimental results demonstrate that IUIENet outperforms the compared methods in both visual quality and quantitative metrics on the UIEB， EUVP， and LSUI benchmark datasets.

Abstract:Due to the complex terrain environment and limited illumination in coal mines， images acquired by video surveillance equipment often have problems such as insufficient brightness， low contrast， color distortion， and loss of detail information. To solve the above problems， a multi-scene low-light image enhancement algorithm of a coal mine based on MCGN （multi-scale calibrated gating network） is proposed. The algorithm is composed of illumination enhancement network， detail enhancement network， color correction network， and gating fusion network. Firstly， the illumination enhancement network estimates the illumination information through a pre-lighting module. On this basis， the illumination enhancement module with spatial enhanced attention is cascading to enhance the capture ability of the occluded area and the local dark area. Subsequently， a self-calibration module is introduced to further improve the overall exposure control ability of the image. Secondly， to preserve and enhance texture and edge details， a multi-level residual structure is designed to form a detail enhancement network， ensuring that important detail information is not lost. Furthermore， in view of the inherent color distortion of the image and the color distortion generated in the enhancement process， a color correction network and a color loss function are constructed. The codec structure is used to decouple the color image into a color histogram， and the natural light color characteristics are learned based on the color histogram to guide the color distribution correction. Finally， in order to realize the organic fusion of the output images of the three networks， a new gating mechanism is designed in the gated fusion network， which learns the optimal fusion weights end-to-end to achieve an effective balance of brightness enhancement， detail restoration， and color correction. Experimental results show that the proposed algorithm is effective in improving image brightness， enriching texture features， and restoring true color. At the same time， the algorithm has good multi-scene applicability and fast reasoning speed， which can meet the actual needs of coal mines and provide strong technical support for coal mine safety production.

Abstract:This paper proposes a fusion algorithm utilizing TransNeXt to address detail loss and artifact generation issues in the fusion of infrared and visible images. Firstly， shallow and deep features are extracted from the source images using convolutional neural networks and TransNeXt. An information compensation module is employed to enhance the semantic information of the infrared shallow features. Secondly， a cross-attention-based fusion module integrates these features， and dynamically adjusts weights based on the importance of different regions in the source images to adapt to scene variations， thereby improving fusion robustness and accuracy. The final fused image is obtained through Transformer-based image reconstruction. In addition，the proposed method constrains the fusion process through a VGG19-based saliency mask loss function， preserving richer information in key regions of the fused results. The experimental results indicate that， compared with the other seven methods， this approach has improved the objective evaluation metrics： namely information entropy， standard deviation， sum of correlation differ-ences， peak signal-to-noise ratio， and pixel feature mutual information， by an average of 10.92%， 14.85%， 24.80%， 2.26%， and 1.30%， respectively. Furthermore， it effectively preserves rich texture information while minimizing artifacts， demonstrating outstanding performance in night light fusion. Additionally， it has achieved superior results in object detection relative to the comparison methods.

Abstract:An improved method called CS-Voxel-RCNN is proposed to address the issue of insufficient detection accuracy of Voxel-RCNN algorithm in detecting small distant targets and occluded targets. Firstly， by introducing three data augmentation methods： random order， random dropout， and random noise， the diversity of training samples is enriched， thereby enhancing the robustness of the model. Secondly， by integrating CBAM in the 2D backbone network and utilizing channel attention mechanism and spatial attention mechanism， multi-scale features are processed in more detail， optimizing the feature fusion effect. Finally， by adding a DIoU loss branch， the original loss function is improved， emphasizing the distance information between the target bounding boxes， thereby improving the accuracy of the target bounding box regression task. Comparative experiments with some classic 3D object detection algorithms on the KITTI dataset are conducted. The results show that the newly proposed algorithm has significantly improved performance， compared with the original Voxel RCNN algorithm， with improvements of 2.91 percentage and 0.87 percentage for pedestrians and cyclists， respectively. The effectiveness of each improvement module is verified through ablation experiments. This series of improvement methods achieve positive results in improving the practicality and accuracy of 3D object detection in real scenes.

Abstract:The inverse problem of electrical impedance tomography （EIT） poses significant challenges due to its seriously non-linear， ill-posed and under-determined nature， which can lead to inaccurate image reconstructions. To address this issue， this paper proposes a novel EIT method based on a multi-mechanism dynamic search. First， the original conductivity distribution matrix of the target region obtained by Tikhonov regularization method is used as the input of the multi-mechanism dynamic search algorithm. Then， the candidate solutions are randomly initialized in the search space， and dynamic optimization of the conductivity distribution is performed based on five selection mechanisms corresponding to population migration and mating behavior. The objective function is then used to calculate the fitness of each individual and the candidate solution with the smallest fitness value is regarded as the optimal solution. Subsequently， the optimal solution is used to compensate the original conductivity distribution， yielding the optimal conductivity distribution. Finally， the imaging quality of this method is verified through simulations and experiments. The results show that the proposed method achieves the lowest root mean square error （RMSE） value， ranging between 0.15 and 0.4， and the highest structural similarity index measure （SSIM） value， varying between 0.55 and 0.85. Compared with other methods， namely LBP， NR， Tikhonov regularization， TV and GA methods， the proposed method demonstrates superior image quality and maintains robust performance under the influence of noise， thereby meeting the requirements for accurate image reconstruction.

Abstract:An improved A* algorithm is proposed to address the issues of tortuous paths， proximity to obstacles， and low search efficiency in global path planning for wheeled inspection robots. First， the cost function of the A* algorithm is optimized to enhance path accuracy. Second， a line-of-sight method incorporating a distance factor is employed to reduce the tortuosity of the global path and increase the safe distance between paths and obstacles. The two-way search strategy is utilized to improve search efficiency. Finally， the global path is smoothed using quasi-uniform B-spline curves. The environment map is modeled using the grid method， and the algorithm is applied to a chlor-alkali chemical inspection robot scenario. Experimental results demonstrate that， compared with the traditional A* algorithm， the improved A* algorithm exhibits superior performance in terms of path safety， and the number of path inflection points. Specifically， the number of path inflection points is reduced by 79.41%. The global path generated by the improved A* algorithm is smoother， better satisfying the requirements for path planning in wheeled inspection robots.

Abstract:Simultaneous localization and mapping （SLAM） technology is widely used in the field of autonomous driving， where accuracy and computational efficiency are the two most important indicators. However， traditional LiDAR odometry faces challenges in accurately and efficiently extracting keyframes， resulting in an excess of redundant frames during map construction. Additionally， the majority of LiDAR odometry systems require aligning each frame to the map， which imposes a substantial computational burden. This paper proposes a dual-mode LiDAR odometry and mapping method based on keyframes. By computing the feature similarity between two point clouds and comparing it with a motion-adaptive threshold， keyframes are extracted. Subsequently， different registration algorithms are applied to keyframes and non-keyframes to minimize computational resource consumption. Furthermore， a weight function is calculated using point horizontal distance information and integrated into the weighted pose constraints. The SLAM system proposed undergoes extensive testing on the KITTI dataset and real vehicles. The results from KITTI sequences 00-10 demonstrate a translational error of only 0.56% and a rotational error of 0.002 1 degree/m. In terms of real-time performance， compared with F-LOAM， our algorithm improves average speed by 26.5%， and even outperforms lightweight system LeGO-LOAM.

Abstract:With the expansion of chip scale and the strengthening of function， the difficulty of verifying chips is increasing geometrically. At present， for the functional coverage of multiple combination incentive cases， the industry’s common practice is to calculate it in the form of fragments or slices according to different use scenarios. This method is easy to operate， but it is difficult to perform a complete coverage analysis of the combination of various configurations under random testing. To solve this problem， a verification method based on a machine learning algorithm for fast convergence and strong universality of coverage is proposed. In this method， each configuration incentive is decomposed according to the weight， and the key cross bins in the function coverage are observed. The data set is collected and trained by the feature that the function point analysis does not consume the simulation time. Through the actual test adjustment， an improved network structure is realized， which can predict the coverage rate of various incentive combinations， and also can pick an incentive input that specifies a coverage threshold. Simulation results show that compared with the random case， the proposed method can significantly reduce the simulation time and effectively reduce the simulation resource occupation. Compared with other network structures， the proposed network achieves faster convergence and higher prediction accuracy.

Abstract:DSPs （digital signal processors） using VLIW （very long instruction word） architecture are widely used in high real-time application scenarios， such as image processing and computer vision. One of the important algorithms in these application areas is the highly parallel multi-directional Sobel algorithm. Implementing and optimizing this algorithm for VLIW DSPs is of great significance. In this paper， we propose a method of optimizing convolutional computation based on VLIW data rearrangement Im2col （image to column） plus matrix multiplication GEMM （general matrix multiplication）， and use DMA （direct memory access） double buffer mechanism to realize the parallelism of data transmission and kernel computation， which reduces the time overhead of waiting for data transmission and the time overhead of kernel computation. The time overhead of waiting for data transmission is reduced， and the multi-directional Sobel algorithm is implemented and optimized on FT-Matrix DSP using this method. The experimental results show that the optimized algorithm achieves 4.96~8.76 times speedup compared with the algorithm in OpenCV image library， and 3.26~6.60 times improvement compared with the TMS320C6678 processor. These results show that the DSP with VLIW architecture has significant advantages in intensive data processing， and the image detection algorithm implemented and optimized on VLIW DSP has a broad application prospect.

Abstract:With the rapid development of the wind power industry， the proportion of wind turbine failures resulting in downtime is also increasing， particularly yaw system failures， which account for nearly one-third （28.7%） of total downtime. To reduce downtime and operational costs， this paper proposes a deep learning model based on SCADA data， named CNN-smart_Linformer （CNN-SLinformer）， for predicting the occurrence time of yaw system failures in wind turbines. This model introduces dynamic self-attention weight calculations for the linear projection matrix， allowing it to adaptively capture changes in the input sequence and significantly enhancing the model’s generalization ability in different operating environments. It combines the advantages of convolutional neural networks （CNN） in local feature extraction with the capability of SLinformer to capture long-term dependencies. Experimental results using actual SCADA data from wind farms show that the CNN-SLinformer model significantly improves prediction accuracy for yaw failure tasks， reducing the score to 144.50， while it also has a shorter runtime， providing an effective predictive tool for wind farms.

Abstract:In the field of missile aerodynamic parameter identification， traditional extended kalman filter （EKF） algorithms often encounter issues such as high computational complexity， low accuracy， and difficulties in solving the system’s Jacobian matrix. To address these challenges， an online identification method for missile aerodynamic parameters based on singular value decomposition-cubature kalman filter （SVD-CKF） is proposed. Leveraging the cubature point linearization characteristic of CKF， this method avoids the direct solution of the Jacobian matrix， thereby reducing computational complexity. Additionally， by introducing Singular Value Decomposition （SVD） technology， it effectively resolves the issue of potential negative definiteness in the covariance matrix that may arise in traditional CKF algorithms， further enhancing filter stability. Simulation results demonstrate that in the context of online identification of aerodynamic parameters for six-degree-of-freedom tactical missiles， the SVD-CKF algorithm exhibits higher identification accuracy， faster convergence speed， and stronger robustness.

Abstract:The effect of data-driven deep learning structural damage identification （SDI） is greatly affected by factors such as structural complexity， model construction method and data size， etc. We introduce graph convolutional neural network （GCN） to integrate the attribute features between structural nodes， explore the complex attribute relationships between nodes from a graph perspective， and provide multi-dimensional learning information for SDI. To this end， a graph convolutional neural network integrating multi-dimensional features of structure （S-GCN） model was designed， which integrates multidimensional structural features. Based on structural vibration data， a damage feature matrix was constructed， and through a derived graph network， the connection relationship between nodes and edges in the graph was represented. An edge index matrix was constructed， and multidimensional feature information such as structural damage status， vibration data， and node attributes was input into GCN for structural damage feature extraction and identification. The application effect of GCN in damage identification driven by multidimensional structural feature information was explored. The feasibility and effectiveness of two steel structure verification methods were verified， and the results showed that S-GCN can integrate multi-dimensional structural feature information， achieve high damage identification accuracy for both structural objects， and demonstrate good noise robustness. Further comparative analysis shows that S-GCN can efficiently update node features and predict node damage status based on inter-node relationships compared with the three non-GCN models. Its damage recognition accuracy， computational efficiency， and network layer evolution process are all better than the comparative models， verifying the effectiveness of integrating structural spatial features in structural damage recognition.

Abstract:A novel spectrum sensing algorithm is proposed for environments characterized by symmetric Alpha-stable （SαS） noise， combining fractional low-order preprocessing with eigenvalue harmonic mean detection. The proposed algorithm employs the ratio of the difference between the maximum eigenvalue and the harmonic average of all eigenvalues to the minimum eigenvalue （DMHMM） as the test statistic. These eigenvalues are calculated from the sample covariance matrix of the received signal， which is preprocessed using fractional lower-order techniques. This algorithm reduces the impact of the non-Gaussian characteristics of SαS noise through fractional low-order operations in the preprocessing stage； and in the detection stage， it uses extreme eigenvalues and eigenvalue harmonic mean to design test statistic. The detection process of the proposed algorithm does not depend on SαS noise parameters and has a wide range of adaptability. On this basis， based on the moment theory of geometric mean of Wishart matrix eigenvalues and the asymptotic distribution theory of maximum and minimum eigenvalues in high-dimensional random matrices， an effective theoretical decision threshold calculation method is proposed for the proposed DMHMM algorithm. This method reduces the complexity of theoretical threshold calculation while improving the reliability of detection results of the primary user signal in SαS noise under non-asymptotic conditions. Monte Carlo simulation results show that the proposed DMHMM algorithm can obtain more reliable decision results than semi-blind DMGM algorithm， and does not require statistical parameters of SαS noise in the detection stage. Due to the comprehensive utilization of the extreme eigenvalues and the harmonic mean of all eigenvalues of the sampled covariance matrix after fractional low order preprocessing， the new algorithm can better reflect the changes in the primary user signal， resulting in high detection probabilities than the traditional MME and CHME algorithms.

Abstract:Aluminum-based composite materials were prepared by a multi-pass equal channel angular pressing （ECAP） process. The microstructure of the composite material was characterized by metallography， scanning electron microscopy， and energy spectrum analysis， and their mechanical properties as well as friction and wear properties of the composite material were tested and analyzed. The research results indicate that SiC particles have good dispersion in the aluminum matrix after three ECAP passes， and the grain size is small. The multi-pass ECAP process can effectively improve the distribution of SiC particles in the aluminum matrix and refine the grain size， thereby simultaneously enhancing the mechanical properties of composite materials. At the same time， under the same friction and wear test conditions， the samples with three ECAP passes show better friction and wear performance than those with a single pass， indicating that increasing the number of processing passes can improve the wear resistance of composite materials.

Abstract:A series of carbon paper-supported spinel structure ferrite MFe2O4 microsphere catalysts （M2+ = Fe2+， Co2+， Ni2+， and Zn2+） were synthesized via an in-situ hydrothermal method using carbon paper as the substrate. The influence of the type of M2+ in the ferrite and the characteristics of an external magnetic field on the catalyst’s oxygen evolution reaction （OER） performance was investigated. Results indicate that carbon paper-supported NiFe2O4 （NFO-Ms/C） exhibits excellent OER performance， with an overpotential of 409 mV at 10 mA·cm?2， a Tafel slope of 78.9 mV·dec?1， and an electrochemically active surface area of 1.6 mF·cm?2. This superior performance is primarily attributed to the abundance of higher-valent M3+ ions in NiFe2O4， coupled with its lower conductivity and abundant oxygen vacancies， which collectively facilitate the OER process. Stability tests reveal that the overpotential of the catalyst increases by only approximately 5% after 60 hours， mainly due to surface reconstruction of the catalyst. An external alternating magnetic field can enhance the OER performance of NFO- Ms/C. When the intensity of the alternating magnetic field is 4.320 mT， the overpotential of NFO-Ms/C decreases from 455 mV to 315 mV at 10 mA·cm?2， representing a 30.8% reduction. This is due to the induced electric field generated by the alternating magnetic field， which increases the concentration of active surface species OH- on the electrode， thereby enhancing the electrode potential， and the magnetocaloric effect provides additional energy to accelerate charge transfer.

Abstract:This paper investigates the effects of TiO2 addition on the composition， microstructure， mechanical properties， and wear resistance of ZTA/ZMn13 composite materials. The results show that the density of ZTA ceramic increases first and then decreases with increasing firing temperature and TiO2 addition. When the firing temperature is 1 600 ℃ and the TiO2 addition is 2 wt.%， the ZTA ceramic has the best comprehensive properties： the relative density reaches the highest value of 98.73%， the flexural strength is （429.60±24.56） MPa， the Vickers hardness （HV） is （1691.73±120.65） N/mm2， the fracture toughness is （7.24±0.38） MPa·m1/2. At 1 450 ℃， the contact angle between ZTA and ZMn13 melt decreases with increasing TiO2 addition， because the TiO2 on the surface of ZTA is connected to the ZMn13 substrate through Ti—O—Fe bonds， thus improving the interfacial wettability between ZTA and ZMn13 substrate. The ZTA ceramic particles prepared as a three-dimensional porous preform are cast at 1 450 ℃ to prepare the ZTA/ZMn13 composite materials. The flexural strength of the material increases first and then decreases with increasing TiO2 content in ZTA， reaching a maximum value of 273.63 MPa when TiO2 content is 2 wt.%. Compared with the ZTA/ZMn13 composite materials prepared without TiO2 （201.72 MPa）， its strength is increased by 35.63%， and its wear resistance is improved by 21.98% under three-body wear conditions.