This pdf file contains the front matter associated with SPIE Proceedings Volume 13249, including the Title Page, Copyright information, Table of Contents, and Conference Committee information

Dietary choices have a substantial impact on the health of an individual. This AI-driven research aims to recognize, classify, and estimate the origin and nutrition of food. The proposed system is trained using a diverse dataset containing images of food (101 Classes) in different lights and environmental conditions. In this research a transfer learning approach applied with ResNet and InceptionV3 architectures using their pre-trained weights with finetuning of hyperparameters (Learning rate, Batch size and Optimizer). As a result of this approach, the ability to learn intricate features relevant to food recognition was retained while training rapidly. The system achieves impressive accuracy: 96.6% and 96.1% respectively for food identification, nutrient, and origin estimation. The system accurately recognizes popular foods like pizza, sushi, and salads, even in low light. Furthermore, to provide reliable food information to end users, we have developed a user-friendly web application. The app allows users to upload pictures of their meals to receive nutritional and origin information, empowering them to make healthier choices. This simplifies the process of making informed dietary choices for individuals.

Depression seriously affects People's Daily life and work, and may even lead to suicide. The method based on deep learn ing is expected to assist the clinical diagnosis of depression more effectively and objectively. Usually, 2D convolutional neural network (CNN) is used for feature extraction for depression level assessment, but this method can only extract stat ic features and will lose dynamic spatiotemporal features. In contrast, 3D CNN model can directly extract temporal and s patial features, which can improve the performance of depression level assessment. In this paper, in order to compare the depression level assessment performance of different 3D convolutional neural networks, we conducted tests using 3D V GGNet, 3D GooleNet, 3D ResNet, 3D SENet and 3D DenseNet networks based on AVEC2013 and AVEC2014 datasets. The experimental results show that the 3D ResNet18 network obtains the best evaluation results,MAE=7.48,RMSE=9.7 0 on the AVEC2013 dataset,MAE= 7.03,RMSE=9.01 on the AVEC2014 dataset. Compared to other existing methods, 3 D ResNet18 shows excellent performance.

Super-resolution image reconstruction is an image processing technology that reconstructs low-resolution images into high-resolution images. At present, there still exist some problems in deep learning algorithm models for super-resolution satellite image reconstruction such as low operating efficiency caused by network overdepth, undesirable information flow between network layers, etc., resulting in low image reconstruction efficiency and insufficient image detail feature extraction. In order to improve the visual effect of satellite image reconstruction and enhance the ability of feature extraction of satellite image, a super-resolution image reconstruction algorithm based on close-cut residual network is proposed. The network model of the algorithm optimizes the ability to extract image feature information, and improves the over-fitting problem caused by the large number of parameters in the coherent framework of the filter. The experimental results show that the close contact clipping residual network can improve the image reconstruction efficiency, the convergence effect is the best, and the visual effect and quantitative results are better than those of the classic deep learning network algorithm. The quantitative indicators of SSIM and PSNR are improved, and the texture details of the visual observation image are clearer.

In the domain of medical image analysis, the privacy of patient data is paramount, yet the need for extensive datasets to train robust models is ever-increasing. This paper introduces a novel dataset distillation approach that leverages multidimensional matching methods, including distribution, gradient, and trajectory matching, to generate synthetic datasets that preserve the utility of original medical datasets while enhancing privacy protection. Applied to the PATHMNIST dataset, a colon pathology benchmark, our method not only achieves superior model performance compared to existing dataset distillation techniques but also significantly improves the privacy of the distilled images, as evidenced by lower L2 norms in pixel-level comparisons. Our findings demonstrate that our approach can serve as a robust framework for generating training-ready datasets that adhere to privacy constraints inherent in medical data applications.

Knowledge graph (KG) is evolving rapidly and play an important role in many applications. Recently, the few-shot knowledge graph completion (FKGC) task, which involves predicting missing information based on a limited number of known facts, has garnered growing interest from both practitioners and researchers. However, these methods often fail to fully utilize neighboring information and overlook the semantic distance between similar entities. To address these problems, we introduce a multi-hop neighbor aggregator based on CNN which designs to make comprehensive use of neighbor information. Additionally, we employ contrastive learning to reduce the semantic distance between similar entities. Compared to the leading baseline GANA, our model shows an improvement of 0.4% and 0.4% on NELL in terms of Hits@1 and Hits@5, respectively, and 5%, 1.4%, and 0.4% on Wiki in terms of Hits@1, Hits@5, and Hits@10. Extensive experiments demonstrate that our method performs exceptionally well on two public datasets.

Robust traffic sign detection and recognition under adverse weather conditions is a critical challenge for autonomous driving systems. This paper presents a novel approach that combines zero-shot learning with contrastive vision-language pre-training to enhance the resilience of traffic sign recognition systems against weather-induced visual impairments. Our method leverages a limited dataset to train a model capable of understanding and processing images degraded by various weather conditions such as rain, fog, and snow without direct exposure to these conditions during training. By integrating descriptive language data with visual cues, our model learns to identify and interpret traffic signs through a generalizable semantic embedding, facilitating robust detection and recognition across unseen weather scenarios. The framework employs a two-stage training process: the initial stage focuses on learning general visual features from minimally weather-affected images, while the subsequent stage enhances the model's ability to predict and adapt to weather-specific distortions using a novel zero-shot learning strategy. Experimental evaluations demonstrate superior performance over traditional methods, particularly in zero-shot scenarios where the model encounters completely novel weather conditions. This approach not only advances the field of image restoration in severe weather but also sets a new standard for the deployment of vision-based systems in real-world environments where variable weather is a common challenge.

This article mainly focuses on the optimization and allocation of aircraft transfer support resources.According to the characteristics of each use and maintenance stage of the aircraft, support resources are divided into five categories: flight support, periodic inspection, troubleshooting inspection, specific task, and field support.The selection principles and determination methods for each type of support resources are provided, providing a new approach and method for the optimal allocation of aircraft transfer support resources. By verifying the accuracy and effectiveness of the method through examples, it can reduce the number and scale of support resources required for aircraft transfer, and achieve the optimal configuration of aircraft transfer support resources.

In recent years, the rapid development of computer vision and artificial intelligence has significantly advanced agricultural applications, particularly in the quality detection and grading of navel oranges. This review explores the latest advancements in machine vision technologies for navel orange quality assessment, focusing on image processing, deep learning algorithms, and multispectral imaging. Image processing techniques, fundamental to machine vision systems, enable the extraction of visual features such as color, shape, and texture, enhancing detection accuracy. Recent studies have demonstrated the efficacy of deep learning algorithms, particularly convolutional neural networks (CNNs), in achieving high-precision grading and real-time performance. Furthermore, multispectral and hyperspectral imaging technologies offer rich spectral information, facilitating more accurate quality detection and maturity assessment. Despite significant progress, challenges such as complex natural environments, lighting conditions, and the high cost of imaging equipment persist. Future research directions include integrating multi-source data fusion, developing efficient deep learning algorithms, and promoting cost-effective imaging technologies. Our proposed advanced algorithm employs the DAHENG machine vision experimental platform and MER-132-30UC camera for high-resolution image acquisition, with HALCON software for sophisticated image processing and real-time classification. This system enhances accuracy and efficiency in navel orange quality detection and grading, addressing key challenges in current agricultural practices.

To address issues of inaccuracy, missed detection, and false detection in license plate recognition algorithms, a deep learning-based recognition network for harsh environments is proposed. The FFA-Net network defogs the original image to enhance clarity. Subsequently, the YOLOX-SD detection network, incorporating the Squeeze-and-Excitation (SE) attention mechanism and deformable convolution, is designed. These enhancements reduce false and missed detections and improve adaptability to geometric deformations, thereby capturing small details on license plates effectively. The detection results are then processed by LPR-Net for information recognition. Experimental results show that the improved YOLOX-SD network achieves 99.69% accuracy, 99.62% recall, and 99.67% average precision, with LPR-Net recognition accuracy reaching 97%. Thus, the YOLOX-SD network algorithm demonstrates significant advantages for license plate detection in adverse weather conditions.

Welding data has a dual dependence on time and space. A welding defect prediction model based on Graph Convolutional Neural Network (GCN) and Long Short Term Memory Network (LSTM) is proposed to address the issue of insufficient spatiotemporal feature extraction in previous welding defect prediction. Taking real-time welding data from automobile factories as the research object, combined with knowledge of welding fragmentation and diversity, a welding defect knowledge graph based on root cause analysis was established for the first time. Then, use graph convolutional neural network GCN to capture the spatial relationships of each input node, and use LSTM to capture the temporal changes of welding data. Fusion features are used to predict welding defects. Compared with classical models such as GCN and LSTM, the proposed GCN-LSTM model improves accuracy and performs better in evaluation metrics such as accuracy, ROC curve, AUC, and recall curve. This study has reference significance for optimizing welding processes and improving welding quality.

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that impacts communication and social interaction, and early detection is vital for timely intervention. This research explores the efficacy of deep learning algorithms in diagnosing ASD using facial expression analysis. We employed several deep learning architectures, including VGG16, VGG19, Xception, and EfficientNetB2, to classify facial expressions in the Autism Image Data datasets sourced from the Kaggle website. Our comparative analysis revealed that the EfficientNetB2 model surpassed the other models in all evaluation metrics: Specificity, Sensitivity, and Accuracy. With a specificity of 90.00% and a sensitivity of 93.33%, EfficientNetB2 demonstrated superior accuracy in identifying ASD cases compared to its counterparts. These results underscore the potential of EfficientNetB2 in enhancing the early diagnostic process for ASD through precise facial expression recognition, suggesting that advanced deep learning networks can significantly aid in the diagnosis and subsequent interventions for individuals with ASD.

High-resolution projection systems often suffer from blurring artifacts that degrade visual quality. To address this challenge, we propose a novel, lightweight method for high-resolution projection deblurring. Our approach involves developing a compact network architecture by replacing standard convolution layers with depthwise separable convolutions. This substitution significantly reduces the model size and computational complexity, making it suitable for resource-constrained devices. Additionally, we integrated a Triplet attention module into the network to enhance crossdimensional feature interactions. This integration enables the model to better capture and utilize cross-dimension information, resulting in improved deblurring performance. Compared to baseline networks using standard convolutions, our method with depthwise separable convolutions and Triplet attention achieves superior deblurring results, as demonstrated by various evaluation metrics.

The point cloud data of city-level transmission lines is huge, and the loading and display speed is slow due to the limitation of storage and computing efficiency. In response to the above problems, this paper provides a three-dimensional visualization method for large-scale point clouds of city-level transmission lines. First, by meshing the large-scale point cloud data, and then using deep learning algorithms and sag algorithms to classify and complete the point cloud data of transmission lines, relatively complete point cloud data of wires and towers are extracted. Finally, combined with the point cloud extraction results, the level of detail construction method based on random sampling is used for three-dimensional visual display. The results show that the method proposed in this paper can improve the quality and visualization efficiency of large-scale point cloud of transmission lines.

Making cultural relics "come alive" has been an important development direction for China's museums and cultural heritage in recent years. At present, with the continuous development of the digital industry, the digital protection and display of museum collections of cultural relics has shown a new development trend, however, some of the cultural relics in the museum due to its own immovability, resulting in difficulties in the establishment of the digital model, the acquisition of high costs etc., to the digitization of cultural relics has brought a series of challenges. Based on these issues, this paper proposes a use of photogrammetry and Reality Capture software to cooperate to establish and improve the virtual three-dimensional model of cultural relics, and then through the form of MySQL and Web to establish a cultural relics digitization platform, the bronze cultural relics in the Hubei Provincial Museum for digitalization and innovation, as a solution to the complexity of building models of the immovable cultural relics in the museum, as well as to achieve the cultural relics model of the Multi-dimensional information presentation, to provide users with a rapid access and browsing methods and platforms. The model reconstruction and digital display of museum artifacts through photogrammetry, which in turn provides new application methods and ideas for the digitization of cultural relics.

Small object detection is a critical task in computer vision, with extensive applications in UAV-based target detection and aerial image analysis. However, current small object detection algorithms often exhibit deficiencies in detection accuracy, leading to frequently missed detections and false positives. To address these challenges, we propose a method for an improved detection model based on YOLOv8s, namely DSD-YOLO. Our contributions are as follows: (a) Replacement of the original feature fusion module with a Convolutional Branch Attention module (C2f_DA), which effectively enhances the model's ability to capture and utilize small object information. (b) Introducing a Small Object Detection Layer (SD_layer) to facilitate multi-scale feature fusion, thereby improving the detection performance for small objects. (c) Incorporating the Dyhead detection head to flexibly capture effective feature information for small objects. Experimental results on the public VisDrone2019 dataset demonstrate our method enhances precision and recall by 6.9% and 8.5%, respectively, with mAP50 and mAP50:95 increases by 9.1% and 6.3%, and detection speed (FPS) increases by 12.6. The enhanced model demonstrates superior performance in detecting small objects in UAV images.

With the rapid development of industrial automation, automated sorting of precision industrial components has become the key to improve production efficiency and reduce costs. In this paper, a machine vision-based sorting method for precision industrial components is proposed, which can effectively recognize and classify iron sheets of different shapes and sizes. This study innovatively integrates image processing, pattern recognition and machine vision technologies to design an efficient automated sorting process. Designed for industrial precision components, this method integrates image preprocessing techniques optimized for industrial-grade image acquisition environments, accurate feature extraction methods for efficient identification of key component attributes, and advanced contour analysis and customized shape matching algorithms to ensure fast and accurate classification of all types of industrial components in automated sorting processes.The experimental results show that the proposed algorithm has significant advantages in improving the sorting accuracy and efficiency, and provides a new gripping and sorting technical solution for the industrial automation field.

The rapid development of science and technology today has made PLC logic controller more mature and perfect, and it is widely used in various fields, including logistics sorting field, which greatly improves the efficiency and accuracy of logistics sorting and promotes the development of logistics enterprises in a more favorable direction. Based on this, this paper optimizes the traditional automated sorting robot action system based on PLC logic controller. Through experimental verification, it is found that the coordinate point error control is lower than 1.0 and the sorting error rate is lower than 5% through optimization, which indicates that the optimized sorting system performs better.

This paper presents an enhanced bidirectional A* algorithm tailored for orchard unmanned vehicles' path planning. It integrates a bidirectional search approach and a turn cost function to bolster safety and efficiency in pathfinding. Notably, the evaluation function undergoes modification to facilitate adaptive adjustments in heuristic term weights, thereby curtailing turning points and path length. To ensure alignment with practical needs, a cubic B-spline curve is employed for path smoothing. Simulation and experimental data substantiate the algorithm's efficacy, revealing a notable 30.08% reduction in navigation distance deviation.

Automatic production is closely linked with industrial robots. However, the precision of robot positioning significantly impacts the stability of the automation system. To address this issue, we establish an industrial automation system comprising an industrial camera, visual processing software, controller, and industrial robot to enable robot self-compensation. The system's visual detection module, comprising visual processing software and an industrial camera, detects offsets and compensates for errors in the industrial robot. This facilitates self-inspection and self-correction of robot grasping errors, enhancing the operational stability of the automatic production system.The experimental results show that this method has strong practicability. Compared with the system without self-compensation, the steady-state error is reduced by 0.01 mm in 1000 times of repeated grabbing, which improves the stability of automatic production and has high popularization and application value.

In order to solve the problem of chaotic program formation in portable bionic robotic arms, which leads to excessive output angle of automation control methods and severe shaking during operation. A portable bionic robotic arm operation path automation control method based on single neuron PID is proposed. This method builds a hierarchical control structure for biomimetic robotic arms, dividing control tasks into multiple levels, and improving the overall flexibility and robustness of the system. At the same time, a single neuron PID algorithm is used to construct a job path controller, which takes the path position data generated by the bionic robotic arm as the control object. By simulating and setting the operation process of the bionic robotic arm, the output function sequence is reasonably arranged to achieve precise control of the end position of the robotic arm. The experimental results show that the proposed method has a control accuracy of no less than 90%, a task completion rate of over 97%, and higher control accuracy and higher operating efficiency of the robotic arm.

With the rapid evolution of science and technology, the imperative for autonomous driving has become ever more pressing. In the domain of self-driving technology, the precise and efficient identification of targets within dynamic driving environments remains a paramount challenge. This paper takes on these challenges directly by introducing a pioneering target detection methodology, building upon an enhanced iteration of YOLOv5. Initially, by substituting FasterNet for the original backbone network, the model leverages its partial convolution features to significantly enhance computational efficiency and inference speed. Additionally, the integration of the SimAM attention mechanism dynamically adjusts the significance of various regions within the feature map, empowering the model to prioritize essential image information and refine detection accuracy. To further bolster the model's accuracy in detecting smaller targets, a dedicated module for small target detection is introduced, meticulously extracting feature map details. Lastly, the incorporation of WIoU as the new loss function mitigates the influence of geometric factors, thereby bolstering the model's robustness. Experimental findings demonstrate that the enhanced algorithm elevates mAP and FPS by 2.9% and 2.6 frames per second, respectively, compared to the original model, effectively addressing the challenges of detection speed and accuracy within autonomous driving scenarios.

To address the issues of low search efficiency and slow convergence in the later stages of traditional ant colony optimization (ACO) algorithms for AGV path planning, this paper proposes an AGV path planning method based on a hybrid particle swarm optimization (PSO) and ant colony optimization (ACO) algorithm. Firstly, a grid map method is used to establish the environmental model. Then, a diversity enhancement mechanism and dynamic adjustment of the inertia factor in the PSO are introduced to perform initial path planning. Based on the initially planned path, pheromone-rich regions are identified on the map. Improvements are made to the pheromone update mechanism and the heuristic function in the node transition probability formula of the ACO to enhance the algorithm's convergence. The improved ACO is then used for path planning. Finally, a secondary optimization is performed on the planned path to obtain the optimal path. The effectiveness of the hybrid algorithm is validated through a comparison with results from a solely improved ACO.

Aiming at the highly nonlinear characteristics of the pneumatic soft finger, the static analysis of its bending performance is carried out, and the multi-objective optimization design is carried out by using the response surface method and the multi-objective genetic algorithm (MOGA). Firstly, the structural design and working principle of the soft finger are described in detail, including its overall structural design, and bending deformation principle based on hyperelastic materials. Then, the structural parameters which have great influence on the bending angle, the maximum stress and the volume are selected as the design variables by the sensitivity analysis method. The experimental design method of BoxBehnken design is used to design the sample points, and the Kriging method is used to construct the approximate model to reveal the relationship between the design variables and the performance index of the soft finger. Combined with the MOGA algorithm, the multi-objective optimization of the structural parameters of the soft finger is realized. After verification, while ensuring the structural strength, the bending angle of the optimal scheme optimization target result is increased by 14.21 % and the volume is reduced by 8.46 %. The results show that the optimization scheme can effectively improve the dynamic and static characteristics of the pneumatic soft finger and provide a theoretical basis for its optimization design.

Studied the dynamic characteristics of a copper anode plate removal robot in three different situations, and conducted mechanical analysis to analyze the dynamic characteristics of the plate removal robot. By establishing a model of the board picking robot and setting parameters in the software, the displacement, velocity, and acceleration curves of the center point at the end of the board picking robot were obtained, and the end effector operated smoothly during the operation process. Obtained displacement, velocity, and acceleration curves of the upper and lower arm joints, with no vibration or shaking at each joint. And the torque diagrams of each joint were obtained, providing a basis for the trajectory design of the subsequent plate taking robot.

An improved Osprey Optimization Algorithm (IOOA) fusing Latin hypercube and Lévy flight is proposed to solve the problem that the original OOA in which the initialized population diversity is low and the iterative search process is unable to jump out of local optima. First, a Latin hypercube initialization strategy is used to generate diverse population individuals. Second, a Lévy flight strategy is used to generate random steps and directions, which enables the population to jump out of the local optimum by random steps in multiple directions during the iteration process. The experimental results are numerically analyzed through simulations on four types of benchmark functions, The IOOA algorithm shows significant improvement in convergence speed and accuracy.

The single-scale filtering method only takes advantage of the fact that it fails to make full use of the characteristics of the observed data at different scales, especially when the system fluctuates drastically, the filtering effect decreases seriously. We propose a novel algorithm of multi-scale fusion and estimation for single sensor target tracking in this paper. Utilizing discrete wavelet transform, we reformulate the state equation and observation equation of Kalman filter in a multi-scale form to found a novel multi-scale Kalman filtering model. By making full use of the signal feature on the diffident scales, the algorithm is more effective for tracking maneuver target, especially in a low signal to noise ratio scenario. A set of Monte Carlo simulations are performed, and the results show the efficiency of the algorithm in this paper.

Aiming at the problem of large degree of core wire offset from the axis of the filter rod in the production process of hard ruffled filter rods, this paper designs an on-line high-speed inspection method based on machine vision for the forming unit of cigarette filter rods, which uses ellipse fitting to identify the filter rods and core wires, and outputs the centre coordinates of the two and determines whether they are qualified or not. The detection speed of the system is fast, high precision, the estimated detection capacity is better than 4200 pcs/min, the detection rate of defective products with core wire offset centre of more than 0.5mm and missing core wire is 99%, which is able to achieve the detection requirements of detecting defective products with core wire filter rods to meet the production needs and guarantee the quality of the filter rods on-line.

As a bionic optimization algorithm, ant colony algorithm searches for the optimal path by simulating the foraging behavior of ants. It has the advantages of strong global search ability and easy implementation, and can be used to solve the path planning problem of patrol robot in campus environment. This paper first constructs the MATLAB simulation model of campus patrol environment, and then designs the path planning method based on ant colony algorithm. The effectiveness and practicability of the algorithm are verified by simulation analysis. The simulation results show that the ant colony algorithm can effectively plan the optimal path for the patrol robot, improve the patrol efficiency and reduce the patrol time. This research provides an efficient and intelligent solution for campus security patrol, and has important practical value

Human-AI interaction is one of the critical factors influencing user behaviors and intentions when using AI fitness coaches to guide exercise. However, it is still unclear how human-AI interactions among AI fitness coaches, users, and human coaches affect users’ behavioral intention. The authors employed an online survey to investigate the factors behind the effect of AI coach-user-human coach interactions on users’ behavioral intention. The results indicated that (1) AI coach-user interaction positively impacts functional value and emotional value; (2) User-user interaction has a favorable influence on both functional and emotional value.; (3) Human coach-user interaction positively influences functional and emotional value; (4) Functional and emotional values favorably affect behavioral intention; (5) Both functional and emotional values act as mediators in the effects of AI coach-user-human coach interactions on users’ behavioral intention. This study offered the theoretical and practical implications of these results.

In this paper, an unattended wind direction adaptive assisted landing device and method are proposed to solve the problem that too much wind or uncertain wind direction leads to too large deviations in the safe landing of UAVs, resulting in limited application scenarios of unattended systems. Through this design, the application environment of the unattended system is increased from level 4 to level 6 in wind speed, and the application scenario is greatly expanded. The design enhances the technical advantages of unattended system products, improves the market competitiveness of unattended system products, and meets the needs of UAV applications in more industries.

The complete scheduling problem of unmanned warehouse AGV is a complex NP-hard problem with a process that includes three parts: task assignment dispatching, path planning, and traffic coordination. This study focuses on determining the optimal task assignment scheme for the AGV considering all known information. With the goal of balancing the workload at each picking station, an unsupervised reinforcement learning-based classification assignment method is proposed. The complex multi-task assignment problem is decomposed into a two-stage assignment problem. In order to reduce the task load of AGV and picking stations, the picking and storing region is first classified, and then tasks are assigned. The algorithm uses a hierarchical reinforcement learning method based on policy gradient to assign the storage nodes by class with the set optimization target. Experiments show that using this approach reduces the difficulty of the AGV scheduling problem and increases the efficiency of the solution.

This study is dedicated to the in-depth exploration and optimization of transmission line recognition technology using the algorithmic framework of YOLOv5. YOLOv5, as an advanced real-time target detection algorithm, possesses excellent speed and accuracy, and thus has great potential for application in the field of transmission line recognition. First, we construct a transmission line recognition model based on YOLOv5. The model makes full use of the efficiency and accuracy of the YOLOv5 algorithm, and realizes the accurate recognition of transmission lines by training and learning from a large number of transmission line images. Second, in terms of practical application, we developed a transmission line recognition system based on the trained model. The system is able to monitor and recognize transmission lines in real time and discover potential safety hazards and faults in time. Through testing and application in real scenarios, we verified the effectiveness and reliability of the system.

As computer vision technology advances, depth estimation has garnered significant interest for addressing depth perception challenges in the field. However, traditional unsupervised monocular infrared image depth estimation methods have difficulty dealing with issues such as edge blurriness and occlusions, resulting in poor depth estimation performance. Therefore, this paper proposes an enhanced DispNet network with local planar guidance layer for unsupervised monocular infrared image depth estimation. Firstly, it addresses the problems of edge blurriness and occluded objects in infrared images through the utilization of multiple small-resolution grayscale blocks and multi-scale feature fusion. Secondly, a plane constraint is introduced on the depth map in the local planar guidance layer to reduce noise and discontinuities in the depth map. On the infrared image dataset Iray, the Abs Rel and RMS log were 0.262 and 0.332, respectively, with an accuracy rate of 94.5% when the threshold metric was less than 1.25³ . The experimental results reveal that the algorithm achieves the expected effect on infrared images, and can provide more reliable visual perception performance when driving at night.

Intelligent identification of fruits and vegetables can not only help reduce waste in picking and selling fruits and vegetables, but also effectively reduce production costs and sales prices. Therefore, it can provide better quality and more affordable fruits and vegetables products. In recent years, with the development of artificial intelligence technology, related algorithms based on deep learning have become a research hotspot. Fruit and vegetable identification technology based on deep learning technology can reduce labor costs in the food industry and standardize the industry. We propose a method for identifying fruits and vegetables based on the EasyDL artificial intelligence development platform. This method uses the EasyDL image classification model to design a data expansion process based on the image characteristics of fruits and vegetables. It utilizes multiple training methods of the EasyDL artificial intelligence development platform. It finally achieves accurate identification of fruit and vegetable types.

To address the issues of insufficient attack capability against black-box face recognition models and poor visual quality of face adversarial samples, we propose the Adv-Makeup++ model. First, we introduce the CBAM attention mechanism module in the encoder. Second, we propose a discriminator momentum enhanced adversarial training strategy, which strengthens the discriminator's robustness to perturbations, thus forcing the generator to produce more powerful adversarial samples. Finally, to overcome the problem of unbalanced adversarial efficiency between the generator and discriminator during GAN training, we introduce a multi-scale feature matching technique to ensure the generator can accurately simulate the feature distribution of real face samples at different scales, thereby achieving a more stable GAN training process. Experiments show that compared to Adv-Makeup, our method significantly improves the attack success rate on the Makeup dataset, while also achieving better visual quality metrics.

During the casting process of copper anode plates, the load is large and the running speed is slow. Therefore, a gear transmission system is used, combined with dual motor drive. However, the tooth side clearance of the transmission system will affect the stability of the casting disc operation. In order to meet the stability requirements of the casting disc, an overall control strategy for dual motor clearance is designed. Design a dual motor clearance synchronous control method based on cross coupling control structure. On the basis of master-slave control, the speed difference of the master-slave motor is added as an influencing factor to the signal input of the slave motor, and a simulation model is established in Simulink. Through comprehensive simulation experiments, it has been verified that the composite clearance control structure can effectively improve the synchronization performance of dual motors.

In this research work, Three non coupled dual servo electric cylinder synchronous motion control schemes based on PLC control have been designed, which are parallel synchronous control, master-slave synchronous control, and virtual axis synchronous control. A dual servo motor synchronous control experimental platform was built, with the core controller being PLC. Three synchronous control schemes were tested under the same load, different loads, and variable loads. Among them, the synchronization performance of dual servo motors under virtual axis synchronization control is better than the other two methods. Then, a dual servo electric cylinder synchronization control experiment was conducted on the virtual axis synchronization control method. The results showed that under this method, the position difference between the two servo electric cylinders was small, and there was no cumulative deviation, indicating good synchronization.

As the largest natural reserve in China, the Three-River-Source Region possesses abundant grassland resources. However, in recent years, the grasslands in this area have been facing severe degradation issues. Overgrazing by livestock, mainly cattle and sheep, who are the primary livestock of the local herdsmen, has led to excessive grassland depletion. Addressing this phenomenon, this paper aims to provide technical support for grassland workers to explore the relationship between livestock grazing and grassland degradation by accurately identifying the number of cattle and sheep. This will enable the formulation of more effective grazing management strategies. This study involved field data collection and manual annotation to construct a dataset comprising 10,000 images of cattle and sheep. Building upon the original YOLOv5 model, optimizations were made by integrating compression techniques and the Squeeze-and-Excitation (SE) module, as well as enhancing the small object detection stage. Following these optimizations, the model achieved performance improvements of 1.2% in precision, 4.03% in recall, and 3.6% in mAP0.5 (mean average precision at an IoU threshold of 0.5) compared to the original model. These enhancements provide efficient and accurate technical means for livestock counting in the Three-River-Source Region.

In complex corn orchard environments, traditional machine vision algorithms often struggle to handle changes in light, occlusion between crops, and interference from weeds, which leads to inaccurate navigation line extraction. This article proposes a corn field navigation route detection method based on improved CenterNet to address this issue. Firstly, we replaced the CenterNet backbone network with ResNeSt, which can effectively improve detection performance when facing small targets and complex hierarchical image content. At the same time, we added CBMA attention mechanism in CenterNet to enhance the recognition of target features of corn seedlings and weaken background interference. We use the midpoint of corn seedlings as the navigation positioning base point, use an improved CenterNet model to detect corn seedlings on both sides, use the center point of corn seedlings as the positioning reference point, and then use the least squares method to fit the center points of corn seedlings on both sides. The experimental results show that the improved CenterNet model has a detection accuracy of 76.44% and a detection speed of 42.09 frames per second. Compared with the original CenterNet model, it has improved by 4.28 percentage points and 1.51 frames per second, respectively. From this, it can be seen that the model can accurately identify corn seedlings, and the average processing time for one frame of image is 0.082 seconds, which fully meets the needs of navigation in corn fields.

The longitudinal control problem of vehicle platoon under disturbance conditions is studied. To eliminate the intervehicle spacing error, a sliding-mode control law with disturbance observer (SMCDO) is proposed for following vehicles in this paper. Firstly, The disturbance is observed through the disturbance observer, and the observation error of the observer converges to zero. Next, a coupling surface is designed for the vehicular platoon, and then, according to Lyapunov stability theorem, the inter-vehicle spacing error of each vehicle is driven to the presented sliding-mode surface and eventually is driven to zero. Furthermore, by choosing appropriate coupling coefficients and utilizing Laplace transforms, the vehicular platoon’s stability is proven. Finally, the simulation demonstrate the effectiveness of the designed SMCDO in handling disturbances, and comparing with the control algorithm without disturbance observer, the superiority of the proposed control algorithm in suppressing disturbance is highlighted.

Firstly, analyze the development of electric vehicle energy storage systems, and then analyze the characteristics of common energy storage systems such as supercapacitors, lithium-ion batteries, and spring elastic energy storage. Use ADVISOR to model the entire electric vehicle and key assemblies, and match the selected energy storage system with the electric vehicle model based on the power requirements of the vehicle. Finally, using ADVISOR, the constructed electric vehicle model is simulated and operated under typical operating conditions, and the operating results are analyzed. Through the study and analysis of the simulation performance comparison of energy storage systems, it was found that composite spring elastic energy storage can increase the range of electric vehicles, demonstrating the obvious advantages of composite spring energy storage systems in electric vehicle design.

In the process of detecting illegal Satellite TV receiver by the unmanned aerial vehicle aerial photography, in order to solve the problems of small size for illegal Satellite TV receiver, low correctness and efficiency of manual identification during the detection process. In this paper, we propose a method of illegal Satellite TV receiver target detection and visualization based on improved YOLO v9. In order to improve the fitting ability of the model, this paper processed the dataset with data enhancement techniques to extend the data. Due to the height of the aerial photography and the angle of the Satellite TV receiver placement, the shape and size of the Satellite TV receiver in the image or video are not fixed, resulting in the accumulation of recognition errors in deep learning networks. In order to solve this problem, this paper incorporates the Alterable Kernel Convolution (AKConv) ideas to improve the algorithm in YOLO v9. After experimental comparison and analysis, the algorithm used in this paper reaches 92.7% in accuracy; mAP_0.5 reaches 97.3%, which are better than other algorithmic networks, and satisfy the accuracy and real-time requirements of illegal Satellite TV receiver detection.

The bearing capacity of the transmission tower can be decreased by the geometric imperfections of chord members. Because the chord members are large and they are welded with foot nails and gusset plates, it is difficult to test their imperfection by the primitive tool such as calipers. Therefore, firstly, the point clouds of the chord members are collected by lase radar; secondly the pipe of the chord member is detected by the deep learning method, then the initial curvature, ellipticity and local unsmooth are calculated; Finally, 5 chord members were tested, and the visual results were given out. The results shows: 1) The ellipticity and local unsmooth of the pipes are not large, except the vicinity of the weld lines; 2) the initial curvature is large, and some pipes has exceed 0.1% length, which may lead to a lower bearing capacity than that calculated by code. Therefore, it is proposed that the bearing capacity can be calculated by the imperfections tested for the important transmission towers