Scene Graph Generation (SGG) serves as an important task in high-level visual understanding, aiming to transform image semantics into structured graph representations, thereby laying the foundation for downstream tasks such as image captioning, visual question answering, and cross-modal retrieval. Compared to traditional two-stage paradigms, one-stage SGG leverages end-to-end training mechanisms to eliminate dependence on external standalone detection modules. For lightweight architectures, it can effectively improve inference efficiency, while complex models with multi-round refinement may not have obvious speed advantages over optimized two-stage methods. Thins paper systematically reviews the technical evolution of one-stage SGG from 2021 to 2026, deeply analyzing the strategies employed by more than 25 works to address core challenges such as combinatorial explosion and long-tailed distribution. Research indicates that although the one-stage paradigm has successfully achieved direct mapping from pixel space to graph structures, the trade-off between accuracy and efficiency, along with the challenge of long-tailed distribution, remains key bottlenecks constraining its development. Looking forward, the integration of large models, efficient inference architectures, and open-world learning will be the core driving forces propelling this technology towards practical applications.

The Infrared-Visible Image Object Detection (IVIOD) algorithm is used to solve the problem of degraded object detection accuracy in complex scenes with a single source by fusing the complementary properties of infrared and visible images. Currently, there are fewer IVIOD researches for unmanned aerial vehicles, and most methods cannot fully consider sensor modal differences and imbalances, causing failure to effectively improve detection accuracy when fusing heterogeneous data. To address these issues, we propose an end-to-end multi-source image object detection framework based on the Decoupled Feature Adaptive Sharing and Cross-Modulation Approach (DFAS-CMA), aiming to improve the accuracy of object detection in infrared and visible images from the viewpoint of unmanned aerial vehicle. DFAS-CMA deploys spatial-channel decoupled re-parametric network to extract decoupled features from different dimensional of multi-source images, effectively capturing small object information and optimizing the feature extraction process. An adaptive feature weight sharing mechanism is then proposed, which effectively mitigates sensor modal differences and imbalances by adaptively fusing heterogeneous data. Ultimately, to improve the robustness of DFAS-CMA in scenarios with small object multi-scale variations, a multi-scale feature cross-modulation aggregation strategy is introduced. Experimental results show that the proposed method improves the mAP0.5 and mAP0.5 & ratio;0.95 by 3.3% and 3.0%, respectively, on the VTUAV-Det dataset. On the DroneVehicle dataset, mAP0.5 and mAP0.5 & ratio;0.95 are improved by 7.1% and 7.0%, respectively, demonstrating the superiority of DFAS-CMA in enhancing the accuracy and generalization ability of multi-source image object detection.

Industrial applications persistently face challenges of data scarcity and high annotation costs, for which few-shot object detection (FSOD) provides an innovative solution. To address the limitations of conventional direct matching methods escalated computational complexity and fine-grained feature loss during aggregation, we propose a novel meta-learning framework with an advanced feature aggregation scheme. Specifically, we first develop a Hierarchical Attention Feature Aggregation (HAFA) module that integrates multi-head attention with a Mixture-of-Experts (MoE) mechanism to achieve precise extraction and decoupled representation of fine-grained features. Building upon this foundation, the Multi-Scale Dynamic Fusion (MSDF) module employs a parallelized architecture incorporating diverse feature interaction operators, coupled with attention-based adaptive dynamic weighting to significantly enhance cross-scale detail feature capture. Comprehensive benchmark evaluations on the PASCAL VOC and MS COCO datasets indicate that our approach achieves state-of-the-art performance across most cases.

Scene Graph Generation is a fundamental task for structured visual understanding, yet existing methods suffer from relation omission or confusion caused by intra-class inconsistency of relations and target occlusion. While recent approaches introduce auxiliary modalities such as word embeddings and knowledge graphs, their static nature restricts adaptability to dynamic relational variations. To address the above issues, we propose DATR, a depth-aware one-stage transformer architecture that leverages geometric hierarchies for fine-grained relation reasoning. DATR comprises three modules: a Cross-Modal Feature Adapter for adaptive RGB-depth fusion to reduce modality heterogeneity; a Bi-modal Guided Encoder that enhances entity discrimination with depth spatial priors; and a Gated-Aware Relation Extraction module that conducts relation classification via adaptive modality weighting. To enable comprehensive evaluation across relation classes, we propose a supercategorylevel metric called S-mR@K. Experiments on the Visual Genome and Open image V6 datasets demonstrate the effectiveness of the proposed framework, highlighting the depth map as an effective geometric prior for SGG.

Object detection, a fundamental computer vision task with broad real-world applications, exhibits vulnerability to adversarial examples comparable to that of classification tasks. The revealed vulnerabilities have become a major bottleneck for its practical deployment. Adversarial Training (AT) is the most effective defense mechanism against such threats. However, existing AT methods for object detection rely on single-source adversarial examples generated from a white-box model, resulting in detectors that remain susceptible to cross-model transferable adversarial perturbations. To address this issue, this study proposed an ensemble adversarial training framework for object detection that leverages multiple models to generate adversarial examples to enrich the training data, coupled with a carefully designed learning strategy. Specifically, we constructed a diverse pool of auxiliary models using publicly available adversarially pretrained backbones, which generate adversarial examples in a round-robin fashion. This operation of enriching adversarial sources helps mitigate architecture-specific overfitting and enhances the black-box defense capability of the detector. In addition, it bypasses the adversarial pretraining step for auxiliary models required in traditional methods, effectively reducing the deployment cost for this computationally intensive task. Furthermore, we introduced a progressive learning strategy to alleviate the white-box performance degradation often associated with ensemble methods. Extensive experiments on the PASCAL VOC and MS-COCO datasets demonstrate that our method is significantly more robust to black-box attacks than single-source AT baselines, achieving performance gains of up to 7%. This work provides valuable insights into promoting the resilience of object detection models against potential real-world attacks. Code and trained checkpoints are available at https://github.com/MonoDone/ODEAT.

In recent years, deep neural networks (DNNs) have achieved widespread success in computer vision tasks, including face recognition, autonomous driving, and medical diagnosis. However, their vulnerability to adversarial attacks has also raised serious concerns regarding system security and reliability. This paper provides a comprehensive overview of recent advances in adversarial defense methods in computer vision. It explores defense strategies for Large Vision-Language Models (LVLMs) while covering traditional adversarial defense methods. The article first explains the basic concepts of adversarial samples, their generation principles, and their actual threats in white-box, black-box and physical worlds. It then systematically examines the multi-dimensional defense strategies based on model architecture design, dynamic adversarial training, and input-output space purification. Meanwhile, for the security challenges of LVLM in the era of significant models, two types of strategies, input-output space defense and dynamic training defense, are discussed. Finally, this paper summarizes the key issues facing adversarial defense and provides an outlook on future research directions.

With the continuous development of intelligent autonomous aerial vehicles (AAV) technology, efficient and accurate sensing of surrounding objects through onboard sensors has become an important research direction. Among these, the object detection is one of the most common perception techniques, but generic object detection methods have low-detection performance in remote sensing images. To address this, we propose the weighted learnable recursive aggregation detection framework, which aims to improve the detection performance in AAV remote sensing images. The network maintains efficiency while ensuring the accuracy of detection. First, to improve the fusion capability of multiscale characterization of small objects, we design a multichannel weights learnable recursive aggregation network. The network improves the multiscale representation fusion capability by dynamically fusing different layers of scale features while recursively aggregating different layers of features. In addition, we design a multichannel recursive residual fusion mechanism for this framework, which is capable of extracting object feature information at different spatial scales and enhances the ability of multiscale characterization of the object. Then, we migrate the multichannel weights learnable recursive aggregation network to different detection frameworks to verify its generalization. Finally, we perform experimental validation using VisDrone 2019, AI-TOD dataset and comparing it with existing methods. The proposed model improves the mAP@0.5 value by 2.1% over the benchmark model on the VisDrone 2019 dataset. On the AI-TOD dataset, our proposed algorithm improves the mAP@0.5 by 1.7% compared with the benchmark model. Also working on the LLVIP dataset, our algorithm achieves 89.8%.

The proliferation of fake news on social media platforms poses serious societal risks, such as eroding public trust, inciting panic, and influencing policy decisions. While automated multimodal fake news detection has emerged as a promising approach, existing methods face three critical limitations: (1) they often fail to capture uncertainty within multimodal data, (2) they struggle with modality heterogeneity, and (3) they lack a balanced focus on both modality-private veracity and cross-modal inconsistencies. In this work, we propose a novel Uncertainty-aware Disentangled Representation Learning (UDRL) framework that addresses these limitations in three key ways. First, our probabilistic representation module models multimodal information as Gaussian distributions, effectively capturing uncertainty and ambiguity. Second, we introduce a disentangled representation learning framework that separates shared and private modality information, enhancing robustness and discrimination. Finally, our uncertainty-aware fusion module dynamically adjusts modality importance based on uncertainty, facilitating more accurate cross-modal interactions. Experimental results on three benchmark datasets demonstrate that UDRL achieves competitive and consistent performance across datasets, validating its effectiveness in multimodal fake news detection.

Camouflaged object detection (COD) is a challenging task that aims to identify objects that blend into their backgrounds. This has various applications in computer vision, including military surveillance, animal behavior analysis, and image editing. However, existing COD methods often face issues of ambiguous boundaries and inaccurate localization due to ineffective feature fusion and boundary guidance. To address these challenges, we propose a new hierarchical synergistic prediction framework, named detail perception synergistic network (DPSNet), which leverages an incremental strategy and a novel detail perception loss (DP Loss) for accurate COD. Specifically, our DPSNet consists of a baseline network and an enhanced network, generating preliminary and fine-grained predictions, respectively. To adapt to the morphological variations of camouflaged objects, we design a morphological awareness adaptation module (MAAM) that adaptively fuses inconsistent morphological and texture features. To enhance the contextual semantic feature representations and obtain clear predictions, we propose a semantic aggregation enhancement module (SAEM) that aggregates contextual semantic features with a deep aggregate group with different receptive fields. Furthermore, to guide the network in learning boundary information for more accurate predictions, we present a DP Loss that balances global information and hard samples. Without any additional tricks, extensive experiments demonstrate that our proposed DPSNet achieves real-time performance and significantly surpasses 37 state-of-the-art (SOTA) methods across four standard metrics on four widely used benchmark datasets. The code will be available at https://github.com/fxle/DPSNet.

In real-world visual recognition tasks, long-tailed distribution is a pervasive challenge, where the extreme class imbalance severely limits the representation learning capability of deep models. Although supervised learning has demonstrated certain potential in long-tailed visual recognition, these models’ gradient updates dominated by head classes often lead to insufficient representation of tail classes, resulting in ambiguous decision boundaries. While existing Supervised Contrastive Learning variants mitigate class bias through instance-level similarity comparison, they are still limited by biased negative sample selection and insufficient modeling of the feature space structure. To address this, we propose Rebalancing Supervised Contrastive Learning (Reb-SupCon), which constructs a balanced and discriminative feature space during model training to alleviate performance deviation. Our method consists of two key components: (1) a dynamic rebalancing factor that automatically adjusts sample contributions through differentiable weighting, thereby establishing class-balanced feature representations; (2) a prototype-aware enhancement module that further improves feature discriminability by explicitly constraining the geometric structure of the feature space through introduced feature prototypes, enabling locally discriminative feature reconstruction. This breaks through the limitations of conventional instance contrastive learning and helps the model to identify more reasonable decision boundaries. Experimental results show that this method demonstrates superior performance on mainstream long-tailed benchmark datasets, with ablation studies and feature visualizations validating the modules’ synergistic effects.

In real-world visual recognition tasks, long-tailed distributions pose a widespread challenge, with extreme class imbalance severely limiting the representational learning capability of deep models. In practice, due to this imbalance, deep models often exhibit poor generalization performance on tail classes. To address this issue, data augmentation through the synthesis of new tail-class samples has become an effective method. One popular approach is CutMix, which explicitly mixes images from tail and other classes, constructing labels based on the ratio of the regions cropped from both images. However, region-based labels completely ignore the inherent semantic information of the augmented samples. To overcome this problem, we propose a saliency-guided local semantic mixing (LSM) method, which uses differentiable block decoupling and semantic-aware local mixing techniques. This method integrates head-class backgrounds while preserving the key discriminative features of tail classes and dynamically assigns labels to effectively augment tail-class samples. This results in efficient balancing of long-tailed data distributions and significant improvements in classification performance. The experimental validation shows that this method demonstrates significant advantages across three long-tailed benchmark datasets, improving classification accuracy by 5.0%, 7.3%, and 6.1%, respectively. Notably, the LSM framework is highly compatible, seamlessly integrating with existing classification models and providing significant performance gains, validating its broad applicability.

Recent advancements in deep learning have driven the rapid proliferation of deepfake generation techniques, raising substantial concerns over digital security and trustworthiness. Most current detection methods primarily focus on spatial or frequency domain features but show limited effectiveness when dealing with compressed videos and cross-dataset scenarios. Observing that mainstream generation methods use frame-by-frame synthesis without adequate temporal consistency constraints, we introduce the Spatiotemporal Attention 3D Network (STA-3D), a novel framework that combines a lightweight spatiotemporal attention module with a 3D convolutional architecture to improve detection robustness. The proposed attention module adopts a symmetric multi-branch architecture, where each branch follows a nearly identical processing pipeline to separately model temporal-channel, temporal-spatial, and intra-spatial correlations. Our framework additionally implements Spatial Pyramid Pooling (SPP) layers along the temporal axis, enabling adaptive modeling regardless of input video length. Furthermore, we mitigate the inherent asymmetry in the quantity of authentic and forged samples by replacing standard cross entropy with focal loss for training. This integration facilitates the simultaneous exploitation of inter-frame temporal discontinuities and intra-frame spatial artifacts, achieving competitive performance across various benchmark datasets under different compression conditions: for the intra-dataset setting on FF++, it improves the average accuracy by 1.09 percentage points compared to existing SOTA, with a more significant gain of 1.63 percentage points under the most challenging C40 compression level (particularly for NeuralTextures, achieving an improvement of 4.05 percentage points); while for the intra-dataset setting, AUC is enhanced by 0.24 percentage points on the DFDC-P dataset.

Adversarial examples are commonly created by solving a constrained optimization problem, typically using sign-based methods like Fast Gradient Sign Method (FGSM). These attacks can benefit from momentum with a constant parameter, such as Momentum Iterative FGSM (MI-FGSM), to enhance black-box transferability. However, the monotonic time-varying momentum parameter is required to guarantee convergence in theory, creating a theory-practice gap. Additionally, recent work shows that sign-based methods fail to converge to the optimum in several convex settings, exacerbating the issue. To address these concerns, we propose a novel method which incorporates both an innovative adaptive momentum parameter without monotonicity assumptions and an adaptive step-size scheme that replaces the sign operation. Furthermore, we derive a regret upper bound for general convex functions. Experiments on multiple models demonstrate the efficacy of our method in generating adversarial examples with human-imperceptible noise while achieving high attack success rates, indicating its superiority over previous adversarial example generation methods.

The increasing difficulty in accurately detecting forged images generated by AIGC(Artificial Intelligence Generative Content) poses many risks, necessitating the development of effective methods to identify and further locate forged areas. In this paper, to facilitate research efforts, we construct a DA-HFNet forged image dataset guided by text or image-assisted GAN and Diffusion model. Our goal is to utilize a hierarchical progressive network to capture forged artifacts at different scales for detection and localization. Specifically, it relies on a dual-attention mechanism to adaptively fuse multi-modal image features in depth, followed by a multi-branch interaction network to thoroughly interact image features at different scales and improve detector performance by leveraging dependencies between layers. Additionally, we extract more sensitive noise fingerprints to obtain more prominent forged artifact features in the forged areas. Extensive experiments validate the effectiveness of our approach, demonstrating significant performance improvements compared to state-of-the-art methods for forged image detection and localization. The code and dataset will be released in the future.

The rapid development of generative technologies has made the production of forged products easier, and AI-generated forged images are increasingly difficult to accurately detect, posing serious privacy risks and cognitive obstacles to individuals and society. Therefore, constructing an effective method that can accurately detect and locate forged regions has become an important task. This paper proposes a hierarchical and progressive forged image detection and localization method called HPUNet. This method assigns more reasonable hierarchical multi-level labels to the dataset as supervisory information at different levels, following cognitive laws. Secondly, multiple types of features are extracted from AI-generated images for detection and localization, and the detection and localization results are combined to enhance the task-relevant features. Subsequently, HPUNet expands the obtained image features into four different resolutions and performs detection and localization at different levels in a coarse-to-fine cognitive order. To address the limited feature field of view caused by inconsistent forgery sizes, we employ three sets of densely cross-connected hierarchical networks for sufficient interaction between feature images at different resolutions. Finally, a UNet network with a soft-threshold-constrained feature enhancement module is used to achieve detection and localization at different scales, and the reliance on a progressive mechanism establishes relationships between different branches. We use ACC and F1 as evaluation metrics, and extensive experiments on our method and the baseline methods demonstrate the effectiveness of our approach.

Multi-modal Entity Alignment (MMEA) aims to establish correlations between modalities such as images and texts to align equivalent entities across different multi-modal knowledge graphs, thereby enhancing knowledge graph coverage and addressing issues of information loss and low coverage in multi-modal knowledge graphs. Existing MMEA techniques mainly focus on heuristic merging paradigms of single-modal embedding. However, due to modality heterogeneity and the absence of visual imagery, current MMEA approaches encounter challenges such as imbalance and ambiguity in multi-modal data fusion, leading to semantic inconsistencies. To address this issue, this paper proposes a feature-balanced Multi-modal Entity Alignment method (FBMEA) and designs a corresponding fusion framework. Different modalities' information is independently encoded, and through adaptive feature fusion and multi-head attention mechanisms, the training effects of weak modalities like visual information are dynamically adjusted to enhance the utilization of long-tail entities. Experimental results on three public bilingual datasets and two cross-graph datasets demonstrate that the model's alignment capability surpasses that of current mainstream models, validating the feasibility and effectiveness of FBMEA.

In addressing the inherent challenge of imbalanced data distribution within long-tailed classification problems and the limitations of traditional classification models in fully extracting intricate information from input images, this paper introduces a pioneering approach: the Tail-Vigor Network (TVNet). Specifically tailored for long-tailed image classification tasks, TVNet aims to explore rare patterns, enhance features, and reveal hidden information. To achieve this, a novel multi-branch network is incorporated within the ResNet architecture to extract and integrate information from various hierarchical levels. Furthermore, this approach involves integrating primitive features from different hierarchical levels into an Adaptive-Context Module (ACM) for feature enhancement, which combines adaptive receptive fields with cross-scale contextual attention mechanisms. Additionally, a novel ReLU activation function is custom-designed to address challenges posed by long-tailed data distributions as it encompasses both linear response stages and nonlinear saturation stages. Empirical validation across multiple public datasets demonstrates that the proposed model yields noteworthy performance enhancements in the realm of long-tailed classification tasks, thus substantiating the efficacy of the proposed approach.

Scene graph generation (SGG) is one of the hottest topics in computer vision and has attracted many interests since it provides rich semantic information between objects. In practice, the SGG datasets are often dual imbalanced, presented as a large number of backgrounds and rarely few foregrounds, and highly skewed foreground relationships categories (i.e., the long-tailed distribution). How to tackle this dual imbalanced problem is crucial but rarely studied in literature. Existing methods only consider the long-tailed distribution of foregrounds classes and ignore the background-foreground imbalance in SGG, which results in a biased model and prevents it from being applied in the downstream tasks widely. To reduce its side effect and make the contributions of different categories equally, we propose a novel debiased SGG method (named DSDI) by incorporating biased resistance loss and causal intervention tree. We first deeply analyze the potential causes of dual imbalanced problem in SGG. Then, to learn more discriminate representation of the foreground by expanding the foreground features space, the biased resistance loss decouples the background classification from foreground relationship recognition. Meanwhile, a causal graph of content and context is designed to remove the context bias and learn unbiased relationship features via casual intervention tree. Extensive experimental results on two extremely imbalanced datasets: VG150 and VrR-VG, demonstrate our DSDI outperforms other state-of-the-art methods. All our models will be available in https://github.com/zhouhao0515/unbiasedSGG-DSDI.

Face forgery represented by DeepFake technique has raised severe societal concerns. Due to the different scales of tampering traces and the different resolutions of face images, adopting common processing pipelines and standard form of convolutional neural networks (CNNs) will lead to problems such as omission, redundancy, and bias when extracting key discriminative features. To solve the above issues, unlike most existing methods that treat face forensics as a vanilla binary classification task, the authors instead reformulate it as a multi-scale object detection problem and propose a novel framework called MSFRNet based on multi-scale feature extraction. Concretely, to alleviate the issues of features omission and redundancy, the authors construct a two-stream prediction network, where the shallow branch discovers small-scale objects such as tiny noise by capturing low-level features with higher resolution and more details, while the deep stream exploits larger receptive fields to detect large-scale blocky artefacts. Moreover, a multi-scale feature extraction module is designed to enrich feature representations in each stream. To solve the problem of features bias and ensure that unbiased feature representations are learned, more appropriate data augmentation approaches are proposed by introducing counterfactual causal reasoning. Extensive experiments demonstrate that our framework outperforms most ordinary binary classifiers and achieves positive performance.