Background Graph representation learning has become a powerful paradigm for modeling the relational structure of visual data in computer vision tasks. Unlike traditional grid-based representations, graph representation learning offers a flexible framework for modeling non-Euclidean structures, which can capture long-range dependencies while facilitating structured reasoning. While several recent surveys review general Graph neural networks (GNNs) architectures and applications across domains, existing literature reviews in computer vision are typically organized around tasks or data modalities. As a result, there remains a lack of a unified perspective for categorizing graph representations by the structural roles they play in visual reasoning. Methodology Following standard literature search guidelines, we collected studies published between 2017 and 2025 from major Computer Vision/Machine Learning conferences and journals using targeted keywords related to graph representation learning. From this corpus , we identified representative works covering seven graph schemas and extracted their graph designs, learning mechanisms, and task-level findings to build our taxonomy. Results Our analysis identifies seven fundamental graph schemas: pixel, point, skeleton, object, region, image, and label graphs, which can form a hierarchy ranging from low-level spatial detail to high-level semantic abstraction. For each schema, we review representative models, discuss graph construction strategies, and compare learning mechanisms, including graph convolutional networks (GCNs), graph attention networks (GATs), spatio-temporal (ST)-GNNs, and graph transformers. Common challenges across schemas include reliance on predefined topologies, uniform neighbor aggregation, difficulty modeling long-range dependencies, and scalability limitations in large graphs. Emerging solutions include graph structure learning, attention-based and adaptive message passing, hierarchical and multiscale graph designs, and integration of graph reasoning into transformer architectures. Conclusions This survey presents a unified taxonomy of graph representation learning in computer vision, synthesizing methodological trends across diverse graph schemas. By clarifying the structural roles of graphs from pixels to labels, we provide a coherent framework for understanding how graph-based reasoning advances visual recognition, geometric understanding, and scene-level interpretation. Traditional challenges, such as adaptive graph construction, efficient reasoning on large graphs, and standardized evaluation, highlight important directions for future research. The insights presented here aim to guide the development of more scalable, interpretable, and context-aware graph-based solutions.
Animal intrusion detection is critical for mitigating agricultural losses, yet real-world deployments remain challenging due to high computational demands and suboptimal hyperparameter configurations. This study presents a robust, multi-class detection framework optimised for real-world agricultural environments. The lightweight SSD Lite MobileNetV3 is benchmarked against the state-of-the-art YOLOv11 family (Nano, Small, and Medium) using a comprehensive, field-collected dataset. To overcome the performance bottlenecks with the default configuration, the framework uses Bayesian Optimisation with Hyperband (BOHB) for compute-efficient hyperparameter tuning. Experimental evaluation shows that BOHB enhances the performance. The optimised models achieved an exceptional F1-score of 0.90 for You Only Look Once version 11 medium (YOLO11m), 0.88 for You Only Look Once version 11 (YOLO11), 0.83 for You Only Look Once version 11 nano (YOLO11n), and 0.68 for Single Shot Detector Lite with MobileNetV3 (SSDLite MobileNetV3). The results show that the proposed framework delivers reliable detection, providing a scalable smart-farming solution to safeguard crops and reduce human-wildlife conflict.
This research presents a detailed framework for predicting traffic speed by integrating spatial and temporal aspects using an advanced hybrid architecture of Graph Sample and Aggregate (GraphSAGE) and Long Short-Term Memory (LSTM) networks, enhanced with transfer learning. Central to this approach is the representation of traffic speed data collected from loop detectors as a graph, with each sensor corresponding to a node. The GraphSAGE convolutions capture complex spatial relationships among nodes, enabling the model to learn how traffic speeds influence one another as a function of geographic proximity and sensor interactions. To capture the dynamic temporal dependencies inherent in traffic behaviours, the LSTM networks are employed, which excel at recognising patterns over time by maintaining information across long sequences. This architecture is particularly effective for predicting traffic speed fluctuations from historical data trends. The framework begins with a pretraining phase on the Performance Measurement System—Bay Area (PEMS-BAY) dataset, which comprises extensive traffic speed records. Subsequently, the model is fine-tuned on the Los Angeles Metropolitan Traffic Dataset (METR-LA) dataset using a limited number of training batches. Limited-batch fine-tuning based on transfer learning is a solution for adapting to data-scarce conditions under the few-shot learning framework that is independent of traditional episodic meta-learning formulations, which require multi-task support-query sets. It is based on the conditions encountered in practice, where the labeled data for the new traffic scenarios is normally very limited. This simulates a few-shot learning scenario, reflecting real-world conditions in which labelled data for new traffic environments is often limited. Through extensive experimentation, the results demonstrate that integrating spatial graph learning via GraphSAGE, temporal modelling via LSTM, and transfer-based limited-batch fine-tuning used here as a practical surrogate for episodic few-shot learning significantly improves traffic speed prediction accuracy, particularly in data-scarce environments. The proposed framework achieves strong performance on the METR-LA benchmark while using only 20 training batches for adaptation, validating its applicability in intelligent traffic management for autonomous vehicles.
Accurate image classification remains challenging when targets share strong visual similarity with surrounding interference, as fine discriminative cues are easily obscured by coupled background and texture noise. Most mainstream attention methods rely on feature reweighting without explicitly modeling feature gaps between targets and interference, leading to unstable discrimination under coupled local-global interference and forming a core bottleneck for classification robustness. To address this limitation, we propose the Grouped Multi-Scale Difference Perception Attention mechanism (GMDA), which shifts the design paradigm from passive weight allocation to active difference amplification. GMDA consists of three synergistic core components. Group-Wise Differential Feature Extraction decouples entangled features via grouped multi-grained extraction to reduce cross-channel interference redundancy. Dual Feature Differential Perception explicitly amplifies subtle discriminative feature differences through dual-path differential operations. Multi-Scale Context Fusion integrates multi-scale contextual information to balance fine-grained local detail perception and global interference suppression. Together they form a progressive processing pipeline that suppresses redundant interference while preserving critical target characteristics, generating precise attention maps and enabling reliable distinction between targets and highly similar backgrounds. We validate GMDA on five representative datasets covering skin lesion classification, fine-grained plant disease recognition, camouflaged object categorization, industrial defect detection and natural scene classification. Extensive experiments are conducted on ResNet-50, MobileNet-V3 and Vision Transformer backbones. We compare GMDA with classic attention methods such as Coordinate Attention and Convolutional Block Attention Module (CBAM), as well as state-of-the-art variants including Coordinate Mixed Pooling Mechanism (CMM) and Adaptive Coordinate Attention (ACA), and supplement the evaluation with systematic component ablation and robustness analyses. Results show that across all test datasets, GMDA delivers an average improvement of over 3% in classification accuracy, macro-F1 and other key metrics relative to mainstream and cutting-edge attention methods, with only marginal increases in parameter count and computational overhead. This work provides an effective solution for image recognition under highly similar interference conditions, and shows promising application potential in agricultural disease detection, medical image analysis, industrial defect inspection and related fields.
Orthogonal time-frequency space (OTFS) modulation is a promising waveform for sixth-generation (6G) wireless systems owing to its robustness in high-mobility and doubly selective channels. However, the inherent two-dimensional symbol spreading and multicarrier synthesis of OTFS generates a high peak-to-average power ratio (PAPR), severely constraining power amplifier efficiency and causing nonlinear distortion and spectral regrowth. This study proposes an Adaptive Scrambling-Based Hadamard Transform (ASB-HT) framework that enables distortionless and side-information-free PAPR reduction through intrinsic data-driven adaptivity. Unlike static transform precoding or probabilistic candidate generation schemes, the proposed method embeds row-wise energy balancing within an orthogonal Hadamard precoder, structurally reducing the likelihood of deterministic peak alignment prior to inverse symplectic finite Fourier transform (ISFFT) spreading. A theoretical analysis establishes the relationship between excess kurtosis and PAPR tail probability, demonstrating that the proposed precoder reduces the fourth-order signal statistics and systematically suppresses the complementary cumulative distribution function (CCDF) of the PAPR. The scheme preserves signal energy, maintains receiver transparency, and introduces only marginal computational overhead. Simulation results across multiple grid sizes (64--512 subcarriers) demonstrate consistent gains, achieving up to 6 dB PAPR reduction at CCDF = 10−3 and up to 7 dB signal-to-noise ratio (SNR) improvement at bit error rate (BER) = 10−3 compared to conventional OTFS, selective mapping (SLM), discrete Fourier transform (DFT)-spread OTFS, genetic algorithm (GA), and deep neural network (DNN)-based approaches. The power spectral density analysis further confirmed the significant suppression of out-of-band emissions, validating the proposed framework for energy-efficient 6G OTFS transceivers.
