Graph neural network (GNN) models are capable of capturing the intrinsic structure and semantic relationships within data and this mechanism grants them substantial potential advantages in the field of computer vision. However, GNN models lack the ability to learn local fine-grained features, which limits their current applications in computer vision. Given the superior capability of convolutional neural networks (CNNs) to learn local detail features, this paper proposes a new module that integrates graph computation layers and convolution computation layers, aiming to learn patch representations that encompass both global and local features, thereby enhancing the performance of GNN models in computer vision tasks. Specifically, (1) the image is divided into patches which are treated as graph nodes, and nodes are connected to their neighbors using KNN, thus representing the image as a graph structure; (2) a module (GraphConv) consisting of a graph network computation layer and a convolution network computation layer is designed, where the former encodes large receptive field information based on the connections between nodes, and the latter learns the fine-grained features within patches using small convolution kernels; (3) the GraphConv module is stacked in a deep model using a residual structure. On the ImageNet-1K dataset and PASCAL VOC object detection dataset, our method significantly improves the performance of the original GNN model. The PyTorch code is available at https://github.com/yuan0038/GraphConvNet.
Although large language models (LLMs) have demonstrated impressive capabilities in generating coherent and fluent text, they often produce irrelevant content when tasked with specialized domains. To address these challenges, Retrieval-Augmented Generation (RAG) combines retrieval and generation processes to enhance the relevance, accuracy, and diversity of LLM responses. However, naive RAG approaches often struggle with precision and recall during the retrieval phase, leading to the selection of misaligned or irrelevant chunks, and sometimes missing critical information. In this work, we propose a novel approach that integrates both word-level and sentence-level retrieval techniques to optimize the retrieval process. By improving the alignment of retrieved information, our method addresses these precision and recall issues more effectively than traditional retrieval approaches. Extensive experiments on benchmark datasets show that our method is significantly superior to existing retrieval strategies, reducing computational overhead and improving accuracy. Our results not only highlight the effectiveness of advanced retrieval strategies in improving LLM performance but also demonstrate their practical implications for scaling NLP systems in real-world applications.
Ensuring safe and accurate drone landings is a critical challenge in the development of autonomous drone systems. This study presents a two-fold approach combining granular semantic segmentation with high-level binary safety classification of aerial imagery. Our method first performs detailed segmentation of environmental elements, then categorizes these segments into "safe to land" and "unsafe to land" zones, providing a more comprehensive and practical solution for autonomous landing decisions. We employed three state-of-the-art deep learning algorithms to perform semantic segmentation on a newly-released dataset of aerial footage from residential neighborhoods. Our models achieved a multiclass Mean Intersection over Union (mIoU) score of 0.82 for detailed segmentation and a binary mIoU score of 0.90 for the critical safe/unsafe landing classification. Among the models tested, U-Net with MobileNetV3 encoder stood out by its better generalization performance, faster training times, and a more compact model size. This study highlights the potential for real-world drone applications, emphasizing its suitability for ensuring safe landings through enhanced environmental mapping and classification.
In recent years, public concern regarding food safety has significantly increased due to the recurrent occurrence of food safety incidents. This paper introduces a comprehensive methodology for assessing food health risks using Fault Tree Analysis (FTA). Initially, FTA is employed to quantify the probability of contamination, thus determining the likelihood of risk occurrence. Subsequently, a diverse set of risk indicators is integrated, and a weighted assessment—incorporating both subjective and objective weighting schemes—is conducted to derive a comprehensive measure of risk severity. The risk matrix approach is then utilized to classify levels of risk likelihood and severity, effectively synthesizing contamination probability with environmental health risk to determine an overall risk value. The proposed framework is applied to evaluate cadmium contamination in rice samples collected in 2019 from the provinces of Hunan, Hubei, and Jiangxi. The findings indicate that Jiangxi represents a high-risk area, Hunan a relatively high-risk area, and Hubei a comparatively lower-risk area.
In this paper, we introduce and experimentally demonstrate a fault detection method for the air conditioner in the EMU train, which usually generates abnormal sounds when the faults occur. The technique utilizes the Support Vector Machine (SVM) to recognize the time-frequency characteristics of the sound generated by the air conditioner, which is extracted using the short-time Fourier transform (STFT). In addition, a detection system is built for the experiment, in which an FPGA and an embedded system are employed for sound acquisition and acoustic signal process, respectively. The penalty factor of the SVM model is optimized and accuracy of 0.977 is achieved. Our method will improve the efficiency and accuracy of fault detection for the air conditioner in the EMU train.
