Heterogeneous Graph Neural Networks (HGNNs), as an advanced method for representation learning on structured data containing multiple types of nodes and edges, have attracted widespread attention. To address the issue of performance degradation with increasing network depth, researchers have introduced the concept of metapaths. However, existing metapath-based methods fall short in capturing neighbor information of different types of nodes, leading to information loss. The selection of metapaths usually relies on domain knowledge and is not easily generalized to unseen graph structures or different tasks. In light of this, this paper proposes an innovative framework-Semantic Structure Encoding Enhanced Heterogeneous Graph Neural Network (SSEH-GNN), which efficiently captures and integrates complex structural and semantic information in the graph through the effective fusion of metapath semantic path information, achieving accurate representation of heterogeneous graph data with a certain degree of generality. To verify the effectiveness of the proposed framework, experiments on node classification tasks were conducted on multiple real-world heterogeneous graph datasets. The experimental results show that SSEH-GNN demonstrates higher predictive accuracy in node classification tasks compared to existing advanced technologies, providing a new methodology and perspective for heterogeneous graph representation learning.
With the rapid development of Internet of Things technology and embedded systems, the application of robotics technology has become an important role in the field of modern technology. However, traditional robot control units often use C language to develop low-level drivers and application programs, which is difficult to develop and has a long cycle, making it a high threshold for non professional developers. Therefore, exploring a robot controller with high compatibility, rich library functions, and high development efficiency is of great significance. MicroPython is a scripting language, and as a lightweight Python implementation, its concise syntax and rich library functions provide new ideas for the development of robot controllers. This article aims to study the design and implementation of a robot controller based on MicroPython. It will comprehensively introduce the research methods and development process of the controller from three aspects: hardware development, firmware design and burning, and image processing applications, providing a new robot controller solution. With the rapid development of Internet of Things technology and embedded systems, the application of robotics technology has become an important role in the field of modern technology. However, traditional robot control units often use C language to develop low-level drivers and application programs, which is difficult to develop and has a long cycle, making it a high threshold for non professional developers. Therefore, exploring a simple, easy-to-use, and highly efficient robot controller is of great significance.
Open world object detection aims to simulate the human process of recognizing objects in real life scenarios. It can identify categories not introduced during the training phase as unknown during the testing phase and recognize these unknowns in subsequent incremental learning. This capability is beyond the reach of traditional object detection, making it closer to real world human experiences and thus garnering significant attention from researchers. However, existing studies largely focus on generating better pseudo labels to improve the recall rate of unknown objects, which evidently neglects the intrinsic differences between objects and backgrounds. Considering important information such as the visual saliency and edge features of objects in realistic scenes, this paper proposes a novel solution called PEC, which directly extracts objects from their background to detect known and unknown objects more accurately. Comprehensive experiments on the M-OWODB benchmark demonstrate that our method outperforms other current mainstream approaches.
Radar can enhance target sensing capability after fusion with visible light to achieve all-weather target detection and identification due to lower requirements for weather and light conditions. However, the mainstream radar and camera fusion methods now use decision-level fusion, which fuses the separately processed radar and image data detection results, and fails to take full advantage of the camera's semantic richness and radar's accurate detection distance. Based on this basic observation, we propose a novel feature-level fusion method, which first optimizes for the camera and radar feature misalignment problem by using a deformable attention mechanism to guide the camera features to offset to the corresponding radar positions and then integrates the optimized camera information into two consecutive cross-attention layers, which incorporate the camera and radar features in turn, exploiting the spatial and contextual relationships to achieve stable and efficient fusion. Extensive experimental results on the popular RADIATE dataset have shown the effectiveness of our method. Compared with the baselines, our method performs better under bad weather conditions. Moreover, the proposed method is robust against various real-world scenes such as rain, fog, and snow.
Many existing studies employ graph neural networks (GNNs) to aggregate information from neighboring nodes and learn node embeddings in a low-dimensional space, aiming to produce layouts that satisfy established aesthetic criteria. However, current GNN approaches for graph layout tasks primarily focus on node features, often neglecting the complex relationships between edge features. To address this limitation and enhance the model's ability to capture complex graph structures, we propose a plug-and-play edge message passing enhancement module that can be easily integrated into existing layout frameworks. This module incorporates both edge distance features and angular relationships between edges, enabling more effective information exchange. We evaluate the usability and effectiveness of this module through quantitative and qualitative experiments.
Defocus blur detection (DBD) aims at accurately extract the in-focus sharp regions in natural images. Despite significant progress, the presence of uniform regions and gradient boundaries remain two of the most challenging problems. In this paper, we use a U-shaped multi-scale image input network for image DBD. The model is a U-shaped encoder-decoder structured network that incorporates inputs from multi-scale images to simultaneously extract features from images of different scales, and a feature fusion module (FFM) of the corresponding scales in the encoder part in order to better learn the features of these different scales. Experiments on two public datasets demonstrate the superior performance of the method in this paper in dealing with defocus blur detection.
Aiming at the poor observation effect of video images due to insufficient light source or unbalanced illumination, and the difficulty of real-time performance of existing algorithms, a brightness equalization algorithm and its parallel optimization method based on layered thinking were proposed. Firstly, the bright and dark areas are divided in the HSV color space. Then the brightness adaptive gain modulation is performed on the low-illumination area. Thirdly, the generated brightness-gain mapping table is applied to the RGB space to achieve low-illumination video image enhancement. For real-time requirements, the 256-level grayscale is mapped to a two-dimensional Luminance Gain Lookup Table (2D-LGLUT) to solve the problem of repeated calculation per frame, and the Intel Advanced Vector Extensions (A VX) instruction set is applied. It realizes efficient and fast processing of 256-bit pixels in a single instruction. The experimental results show that the adopted method achieves the real-time $(\geq 7\text{lf}/\mathrm{s})$ processing target of full High Definition (HD, 1920xl080) video images while ensuring the brightness equalization effect.
Autoimmune encephalitis (AE), particularly Leucine-rich Glioma-Inactivated-1 autoimmune encephalitis (LGII-AE), presents complex diagnostic challenges. The brain regions and extracted features used in previous machine learning works are not fine-grained. This study aims to bridge these gaps by enhancing the granularity of the radiomic analysis, and we encoded a total of 165 brain regions using 107 radiomic features. To further investigate the pathogenesis of facio brachial dystonic seizures (FBDS), a typical clinical symptoms of LGII-AE, three tasks are designed: identifying CDLGII-AE, @FBDS subgroup and @non-FBDS subgroup from healthy control (HC) group. We employed the Mann-Whitney U-test and a Random Forest (RF) algorithm to evaluate feature significance comprehensively. Furthermore, we proposed a feature selection and ensemble learning method for classification, achieving consistency ratios with the average standardized uptake value ratio (SUVr) results of 85 %, 60 %, and 80 % respectively. Our findings indicate that morphological and first-order features are significant biomarkers. The temporal pole middle temporal gyrus_R, inferior temporal gyrus_R, middle temporal gyrus_L and superior frontal gyrus,dorsolateral_R may be linked to LGII-AE. Additionally, certain areas in the cerebellum, frontal and hippocampus regions could be crucial in FBDS. Our study provides deeper insights into LGII-AE's pathogenesis and offers a more comprehensive framework for future research into its diagnosis.
Accessing complex medical data, especially temporal information, presents a significant challenge for non-technical users, including healthcare professionals not versed in technology or query languages like SPARQL. This paper introduces an enhanced form-based SPARQL query builder that offers effortless access to temporal medical data. The described system simplifies constructing SPARQL queries into an intuitive, form-driven interface that abstracts and hides the technical complexities. Novel enhancement is primarily the user's ability to filter and order data with column constraints. By focusing on usability, the system empowers users to retrieve temporal medical information without needing in-depth knowledge of database systems or SPARQL syntax. We evaluate the system's usability and performance by testing it with non-technical healthcare professionals. Results demonstrate that users can easily navigate the interface, significantly reduce the time spent querying the database, and effectively retrieve data. This approach benefits from FAIR data principles and enhances data reusability in healthcare settings for reaching accuracy and precision for follow-up analysis in evidence-based research, facilitating better decision-making that can improve future patient outcomes.
Planktonic algae are widely existed biological resources in nature, and realizing real-time accurate and intelligent detection of planktonic algae is of great significance in many fields. Whole-slide scanning technology is a fundamental method to acquire and digitize microscopic images of planktonic algae, making constituting a dataset for deep learning possible. Currently, deep learning-based intelligent identification of planktonic algae is limited by the scarcity of large-scale multi-class planktonic algae dataset and the difficulty of extracting features among its various classes, which makes it hard to the detection accuracy and identification types at the same time. To solve the above problems, we propose a YOLO based method for planktonic algae recognition and analysis. In order to solve the problem of very limited samples in the dataset, we combine the mosaic and traditional data augmentation to expand the size of the dataset to ten times of the original one. To mitigate the impact of scale variations among different algal species on the proposed method, we combine the Gather-and-Distribute mechanism with the Attention Scale Sequence Fusion module, named as YOLOv8-GA, to comprehensively learning multi-scale features. We conduct experiments on our established planktonic algae dataset with 8 types, and the results show that the proposed method can get higher recognition accuracy. The accuracy rate of YOLOv8-GA for the recognition of eight types of planktonic algae is 90.1%, and the mAP@50 can be up to 92%, Compared to the unenhanced model without data augmentation, the accuracy has increased by 13.3%, while the mAP@50 has improved by 4.3%.
ResNet effectively addresses the vanishing gradient problem in deep networks by introducing skip connections, making it a leading model in image classification tasks. This paper proposes an improved ResNet-based model leveraging transfer learning from ImageNet, initially incorporating additional layers such as Flatten, fully connected layers, L2 regularization, Batch Normalization, and Dropout. The research began with an “initial improved version” of ResNet50, designed to enhance stability and generalization. Subsequent experiments focused on further optimizations inspired by recent studies, particularly the strategic placement of Dropout layers. Comparative experiments on the CIFAR-10 dataset included the original ResNet50, transfer learning-enhanced ResNet, and multiple improved variants. The results demonstrate that the initial improved model significantly enhances stability and generalization compared to the original ResNet. Further optimizations, such as placing Dropout before the Global Average Pooling (GAP) layer, yielded the best performance on the training set, while placing Dropout after the Flatten layer and before the Dense (128) layer resulted in superior validation performance. This study underscores the benefits of a stepwise optimization approach, combining transfer learning, structural modifications, and strategic Dropout placement to enhance image classification performance.
Damage assessment has become one of the important links of precision operations in modern war. With the development of reconnaissance platform technology and the improvement of image resolution, image-based damage assessment is favored by modern war because of its intuitiveness and strong anti-falsification. In this paper, the complete flow of damage assessment based on image is designed, and a damage assessment decision model based on proposed damage assessment indicators is constructed for complex image. The model is used to assessment the damage information extracted by different methods. Experiments show that the constructed assessment indicators and assessment model are effective in evaluating the degree of damage, which consistent with visual evaluation results. Experiments also show that damage region extraction and damage assessment model are very important for accurately evaluating the results of the damage.
Glioma is the most common and aggressive brain tumor leading to a very short life expectancy in their highest grade. Accurate segmentation of multiple gliomas from multimodal MRI is a prerequisite for medical diagnosis and contributes to improving the quality of life of oncological patients. In order to effectively extract the characteristics of glioma and improve the segmentation accuracy, we propose a hierarchical segmentation algorithm based on superpixel sclustering for HGGs. The MRI image is enhanced and denoised with preprocessing methods. Then the MRI image is segmented with superpixel approaches. The dynamic region merging algorithm is performed to progressively integrate the neighbouring superpixels, from which the tumor region is extracted with random forest. Finally, three kinds of lesion regions, necrosis, oedema, and enhancing tissues are detected with the K-means clustering method. Experimental results show that the proposed algorithm has an improvement in segmentation accuracy.
In this paper, we propose a multi-dimensional insider threat detection framework (MITD), which aims to improve the accuracy of detection by comprehensively analyzing user behavior patterns. We combine statistical features and behavioral sequences and use sequence pattern recognition, graph structure information mining, and adaptive feature extraction methods. It can reveal user behavior from multiple dimensions. Then, we intelligently fuse the extracted multi-dimensional features to generate a comprehensive feature representation. It is further fused with statistical features through two strategies: concat and bitwise addition. Experimental results based on the CERT r4.2 dataset show that the MITD framework significantly improves in terms of accuracy, precision, recall, and F1 score, reaching 93.2% accuracy and 93.5% recall. The results show that the MITD framework effectively improves the performance of insider threat detection and has great potential in practical applications.
Brain-controlled system has been used in assisting disabilities, and how to enable such groups to have the ability to communicate with the surroundings has become the focus of research. However, due to the individual differences and the dynamic time-varying brain states, the fixed time window used in the traditional Filter Bank Canonical Correlation Analysis (FBCCA) algorithm for Steady-State Visual Evoked Potential-based Brain-Computer Interface (SSVEP-BCI) is not optimal, which limits its generalizability. Therefore, in this paper, based on the SSVEP experimental paradigm, we propose an improved FBCCA algorithm with dynamic window based on the difference level between correlation coefficients. Compared with the traditional FBCCA algorithm, the accuracy is significantly improved. Furthermore, this paper designs a BCI-based robotic arm control system, achieving real-time and precise object grasping. This validates the feasibility and effectiveness of the proposed algorithm. Our findings may provide a theoretical basis and technical support for the practical application of BCI.
Particle size distribution or aggregate gradation is the key to ensure the quality of concrete and optimize the structure design. In this paper, a method of measuring aggregate particle size distribution based on smartphone image is proposed. Firstly, the image contrast is enhanced by gray-scale and luminance transform, and the noise is removed by adaptive median filtering algorithm, and the edge of particles is preserved. Then the watershed algorithm is used to segment the image on the basis of clustering, and the two-dimensional shape and size parameters are extracted from the segmentation results. Compared with the screening method and the manual measurement method, the cumulative error rate of this method is controlled within 5%, which verifies its reliability. This method is suitable for rapid assessment of aggregate particle size distribution and provides effective technical support for quality control of concrete engineering.
In order to address the issue of incomplete feature extraction in traditional oral fluency scoring methods, a method based on feature fusion is proposed in the paper, which integrates traditional acoustic features and deep features of audio to complete the scoring of oral fluency. For the audio to be tested, IS10_paraling features are extracted as traditional acoustic features, and a fine-tuned wav2vec2.0 pre-trained model is used to extract deep features of the audio. These two types of features are concatenated to form fusion features, which are then input into a multilayer perceptron scoring model to obtain the oral fluency score. Experimental results on the speechocean762 dataset show that compared to scoring methods based on single types of audio features, the fusion feature-based method improves scoring performance. The multilayer perceptron scoring model based on fusion features achieves a Pearson correlation coefficient of 0.744 with human scoring, demonstrating the effectiveness of the proposed method.
Most of the existing ensemble clustering algorithms improve the performance by weighting the basic clusters to reduce the influence of low-quality basic clusters on the final clustering results. Low-quality base clustering can be understood as misclassifying sample points, which are presented as discrete points in the co-association matrix. So, we proposed a multi-view ensemble clustering algorithm based on the weighting of kernel density estimation, starting from the density distribution of discrete points in the co-association matrix. Firstly, the different sets of basic clusters are made into a co-association matrix using the evidence accumulation model. Secondly, we converted the co-association matrix into a sparse matrix, and then we calculated the density distribution weights of the nonzero elements in the matrix using kernel density estimation. Next we let these weights multiply with the sparse matrix. Thirdly, we reduce the weighted sparse matrix to the shape of the initial co-association matrix to obtain the density-weighted co-association matrix. Next, we use the K-means clustering algorithm on the weighted co-association matrix to obtain the final clustering results. Finally, the algorithm is subjected to comparison experiments and ablation experiments on five commonly used datasets. The experimental results show that the proposed ensemble clustering algorithm with kernel density estimation weighting performs better than other comparative algorithms.
This paper proposes a novel approach to optimize the plant recognition function of intelligent agricultural irrigation system, utilizing Stable Diffusion model and Lora (Low-Rank Adaptation) model. Stable Diffusion model is a latent diffusion model that converts text into images, while Lora is a method of fine-tuning large models by reducing trainable parameters while maintaining model performance as much as possible. By combining the Stable Diffusion model and Lora model, this paper trains custom models for each plant species using a small amount of data and generates corresponding images of each variety to assist training, thus optimizing classification results. In addition, to address the issue of imbalanced data, this study also introduces Focal Loss to mitigate the impact of both class imbalance and difficulty in classification. Building upon Focal Loss, CFloss is created as the model's loss function. This method was tested on the OxfordFlower102 dataset used by the original model, and the results demonstrate that it effectively improves classification accuracy and significantly enhances the plant recognition function of the irrigation system.
Super-resolution is of great importance when it comes to hyperspectral image processing. The existing methods have not explored the frequency component relationships between low-resolution hyperspectral images (LR-HSIs), high-resolution multispectral images (HR-MSIs), and high-resolution hyperspectral images (HR-HSIs). Drawing from empirical observations of the correlation between amplitude and the illumination / spectral characteristics of images, as well as the correlation between phase and the structural aspects of images, we show that the amplitude of LR-HSI is closer to HR-HSI, and the phase of HR-MSI is closer to HR-HSI. Based on this, we devise the Phase-amplitude Decoupled Modulator (PaDM), which extracts amplitude information from LR-HSI and phase information from HR-MSI to regulate the hyperspectral image super-resolution (HSISR) process. Experiments demonstrate the effectiveness of the proposed method.
