Multi-view pedestrian detection has been widely used in video surveillance and autonomous driving, but existing multi-view approaches often exhibit performance degradation when deployed in scenarios significantly different from their training conditions. Such generalisation issues are particularly pronounced in situations where models are trained in daytime scenarios and tested in nighttime scenarios, or in cross-dataset settings. This degradation is primarily caused by substantial differences in appearances. In this paper, we propose MVFDA, a frequency domain augmentation method tailored for multi-view pedestrian detection that improves generalisation by perturbing appearance without disrupting geometric consistency. MVFDA builds on two core innovations. First, intra-view low-frequency amplitude swapping exchanges low-frequency components of the Fourier amplitude spectrum only between training images from the same camera view, without use of target-domain data, thereby preserving the geometric consistency essential for multi-view fusion. Second, a MaxDiff strategy automatically selects the reference image that exhibits the largest appearance difference from each input, encouraging each swap operation to generate meaningful diversity. Comprehensive evaluations on the Wildtrack, MultiviewX and MVPerception (Day-to-Night) datasets, as well as in cross-dataset scenarios, demonstrate that MVFDA augmented detectors achieve state-of-the-art generalisation performance in domain shift scenarios while maintaining competitive performance in standard scenarios.
In the context of complex traffic scenes, accident detection plays a vital role in maintaining road safety and reducing response times in critical scenarios. The challenge is much greater, however, for night time conditions, where illumination is poor, and motion blur is happening. This paper introduces a unified framework for robust nighttime accident detection, based on both attention and proximity. The model uses channel attention mechanisms with a hierarchical Swin Transformer backbone to extract discriminative and illumination-invariant spatial features. To enhance the robustness of the classification, distribution-aware proximity classification is adopted to measure the similarity to class distributions in feature space, which in turn helps improve the separation between normal and accident samples. In order to mitigate the impact of data imbalance, focal loss is used when training the model. Experimental evaluation on the NiAD-Large dataset demonstrates that the proposed framework improves accuracy from 76.40
In industrial applications, AI-based methods are gaining increasing significance in optical systems for tasks such as identification, inspection, and classification. The appeal of these methods lies in their operator-friendly implementation and often superior performance, particularly in complex classification tasks. However, despite the substantial potential of AI-based methods, their application is frequently constrained by the significant resources needed to acquire an appropriate training dataset. For instance, in the sand casting industry, the complexity of optical inspection for cast parts is compounded by considerable variation in local surface topography and global object geometry. This challenge pertains not only to the volume of images required but also to the extensive process of accurate labelling. In this work, we investigate the effectiveness and limitations of synthetic training data for an AI-based optical code reader, designed for the identification and tracking of cast parts. This scanner is capable of detecting and classifying a unique code, called Cast Code, developed specifically for the casting industry, enabling the identification of individual cast part numbers. For synthetic image generation, we utilize physically based rendering, which allows comprehensive control over all rendering parameters. This approach facilitates both a systematic investigation of parameter relevance and an automated labelling process for the training datasets. Our results indicate that detailed geometric modelling of both, the local surface topography and the global object geometry of the pins has a notable positive impact on the neural network’s recognition accuracy, achieving accuracy rates of up to 84
Timely detection and quantification of plant abnormalities in agricultural fields is critical, as delays can significantly impact crop health and yield. Automating this process can offer substantial economic and environmental benefits. The focus of this work is on detecting and quantifying abnormalities in maize fields by analyzing high-resolution RGB images captured by Unmanned Aerial Vehicles (UAVs). Common approaches rely on limited training data that are also tuned to specific field characteristics. As a result, models trained on one field do not perform well on data from another field due to the domain shift. To address this challenge, we present a deep learning framework for robust and accurate detection and quantification of abnormalities in maize fields, independently of growth stage, abnormality severity, or soil type. Central to the proposed approach is the idea of standardizing data collected under diverse field conditions by mapping them to a shared domain using unsupervised domain adaptation (UDA). UDA preserves critical abnormality-related information while filtering out irrelevant, field-specific variations. To further enhance data quality, we filter out poorly adapted images that introduce artifacts using a methodology we refer to as selective UDA. We leverage the shared domain to enhance training by performing data augmentation and improve domain generalization by bridging the domain shift between training and test images. During training, data collected under various field conditions are effectively aggregated by mapping them into a shared domain, enriching the training data with domain-adapted images rather than synthetic data alone. During testing, incoming images are mapped into the same shared domain. Our framework employs Vision Transformers for detection and quantification, along with an ensemble of CycleGANs for domain adaptation. We demonstrate the effectiveness of the proposed approach on a publicly available dataset containing UAV-collected RGB images of healthy and unhealthy maize plants of various growth stages from two different maize fields exhibiting variable conditions.
End-to-end scene text spotting jointly localizes and recognizes text within a unified framework. Explicit-point methods such as DeepSolo provide an effective representation of arbitrary text geometry, but their recognition remains largely driven by local visual evidence and lacks explicit contextual support when individual character observations are ambiguous. To address this limitation, we propose DeepSolo-Sem, a semantically adaptive extension of explicit-point text spotting that introduces character-sequence context while retaining the original geometry pathway. Specifically, we develop a Visual-Semantic Alignment Module (VSAM) that converts detached soft-character posteriors at ordered points into continuous expected character embeddings, encodes the resulting sequence with a compact denoising-pretrained frozen bidirectional gated recurrent unit (BiGRU), and projects the contextual states into semantic point features. We further introduce an Adaptive Multi-modal Gated Fusion Network (AMGFN), in which within-instance cross-attention first produces a semantic-conditioned refinement of each visual point feature, and an uncertainty-conditioned channel-wise gate controls the strength of that refinement before final recognition. This design enables contextual reasoning directly over the structured outputs of the explicit-point decoder without relying on discrete transcript correction or a large token-level language model. Extensive experiments on Total-Text, ICDAR 2015, and CTW1500 demonstrate consistent improvements over the reproduced DeepSolo baseline and competitive performance against recent scene text spotters. Component ablations and semantic-encoder comparisons support the proposed alignment and adaptive fusion mechanisms, and the gate-behavior analysis is consistent with the intended uncertainty response. The compact character-sequence prior introduces only modest additional model and inference overhead.
Accurate polyp segmentation is essential for early colorectal cancer diagnosis but remains challenging due to expensive pixel-level annotations, significant variations in polyp size, blurred boundaries, and multiple polyps. To tackle these issues, we propose a weakly supervised segmentation framework using only bounding box annotations. The framework integrates bounding-box priors with MedSAM semantic predictions to generate reliable foreground-background labels. An adaptive prototype generation and fusion strategy is introduced to construct multidimensional foreground and background prototypes, improving discrimination of diverse polyp morphologies. The network is jointly optimized with boundary alignment, continuity, and adaptive correlation consistency losses, enhancing boundary accuracy and regional coherence. Experiments on five public datasets show that the proposed method achieves competitive performance under bounding box supervision, outperforming existing box-supervised approaches and approaching the performance of several fully supervised methods on multiple datasets and evaluation metrics.
Image tampering localization is an important task in digital forensics that aims to identify manipulated regions at the pixel level. Existing methods may be insufficiently sensitive to subtle high-frequency tampering traces, may not fully exploit frequency-aware information during feature fusion, and may have difficulty emphasizing suspicious regions under diverse image contents. To address these issues, this paper presents a high-frequency-aware and dynamic focusing network, termed the Frequency-Aware Dynamic Network (FAD-Net), for image tampering localization. FAD-Net adopts ConvNeXt-B within an encoder–decoder architecture and combines multi-level feature enhancement with localization-oriented feature refinement. Specifically, the Frequency Focus Module (FFM) is embedded in the lateral connections of the first three scales to incorporate frequency-aware processing into multi-scale feature fusion. Through its dual-branch spatial-channel design, FFM separates and emphasizes high-frequency components in the frequency domain, with the aim of preserving subtle forensic cues, such as edge artifacts and noise inconsistencies, during feature aggregation. A Pyramid Pooling Module (PPM) is introduced at the highest feature level to enlarge the receptive field and capture multi-scale contextual information. In addition, the Dynamic Channel-Spatial Attention Module (DyCSAM) is placed in the multi-scale feature fusion path to refine localization features through the coordination of channel attention and dynamic convolutional kernels. Experiments are conducted on four public datasets, namely Columbia, CASIAv1, NIST16, and IMD2020. The results show that FAD-Net achieves competitive F1 and IoU performance across the evaluated datasets, including post-processed scenarios. Overall, FAD-Net provides a task-oriented integration of frequency-aware feature fusion and dynamic feature refinement for image tampering localization.
Sparse-view computed tomography (CT) reconstruction is an important approach to reducing harmful X-ray exposure while maintaining image quality. Conventional CNN-based methods effectively extract local features but struggle to capture long-range dependencies. Transformer-based models address this limitation by modeling global relationships across tomographic regions; however, they typically use large image tokens (such as 16 × 16 pixels), causing locality bias and limiting the modeling of fine structural details. Since abnormal CT regions are often small, tokens should ideally be much finer (for example, 2 × 2 pixels). This greatly increases the number of tokens, leading to the quadratic complexity 𝒪(N^2) in standard self-attention. To overcome this challenge, recent work has explored efficient approximations of self-attention. Motivated by these advances, we investigate treating the regularization gradient of CT images as a long token sequence, where each token corresponds to a 2 × 2 pixel region. This representation enables efficient modeling of long-range dependencies using modern sequence-processing architectures. We integrate selective State Space Model mechanisms from Mamba and three approximate self-attention methods, including Longformer, LongNet, and Nyströmformer, into a regularization-based reconstruction network. Experiments on the AAPM and DeepLesion public datasets show that Nyström attention achieves SSIM scores of 0.9607 and 0.9711 , outperforming competitive state-of-the-art methods by 1.3% and 3.1% . While conventional reconstruction methods perform well under low-noise conditions but deteriorate in noisy settings, our Nyström-based network maintains strong stability and reconstruction quality. Overall, this study demonstrates that representing CT images with small pixel-level tokens provides an effective and computationally efficient way to capture long-range dependencies for sparse-view CT reconstruction.
Gray-level co-occurrence matrix features are widely applicable to texture analysis and texture-related tasks, but their discriminative power is reduced to varying degrees in complex backgrounds. To overcome this limitation, we propose a Redundant GLCM (RGLCM) framework that leverages multi-scale image decomposition and a hybrid feature extraction scheme. The input image is first processed using multiple Generalized Gaussian Filters with distinct shape parameters (β) to generate low-frequency structural components, from which corresponding high-frequency difference maps are computed. Rather than symmetrically extracting high-overhead descriptors across all layers, the framework applies a structured mechanism: five rotation-invariant GLCM features are systematically extracted from both the n low-frequency filtered components and n high-frequency difference maps to isolate localized anomalies and background regularities simultaneously, yielding a comprehensive 2n × 5 dimensional representation. A data-driven selection strategy based on feature importance extracted with Random Forest is then employed to eliminate redundancy and identify the most discriminative descriptors. Experimental evaluation on the AITEX textile defect dataset demonstrates that the proposed RGLCM methodology significantly outperforms conventional single-scale methods in both detection accuracy and robustness while maintaining a highly compact feature representation. This method provides a structurally rigorous and computationally efficient framework for fabric detection, which can accurately capture microscopic anomalies and global texture patterns, and reduce feature redundancy.
As industries increasingly move towards automation, the ability to visually localize objects and accurately estimate their pose, including rigid and non-rigid transformations, becomes critical. Conventional pose estimation methods are constrained to specific object categories, whereas category-agnostic approaches remain limited in accuracy and in their ability to handle composite objects with repeated and non-unique components. In this paper, we present a general framework for object localization based on keypoint graphs, where nodes represent keypoints and edges encode pairwise relations. Both nodes and edges are annotated with natural language descriptions. We introduce a novel neural network architecture that accepts natural language-labeled keypoint graphs as prompts and predicts the image coordinates of graph nodes. Furthermore, we make several industrial datasets publicly available and demonstrate that our method substantially outperforms existing methods. Additionally, we achieve strong performance on the public MP-100 benchmark while offering greater flexibility in representing and localizing complex objects.
The non-invasive nature of facial authentication makes it a preferred biometric security method as it allows remote identity verification through contactless authentication. The security of facial authentication systems faces major threats because of Presentation Attacks. The proposed method implements a video-based Presentation Attack Detection through both intra-dataset and inter-dataset evaluation methods. It implements a CNN framework which combines attention mechanisms with LSTM layers to extract both spatial and temporal information from data. The proposed method is evaluated through three established benchmark datasets including CASIA-FASD, Replay-Attack and Print-Attack while comparing its results to current deep learning techniques and SOTA solutions. The experimental results demonstrate that the proposed framework outperforms existing methods in intra and cross-dataset evaluation while maintaining a compact model size. The model achieved low HTER values of 0.16 for the CASIA-to-Replay protocol and 0.12 for the Replay-to-CASIA protocol. These findings indicate the strong generalization capability of the proposed approach across different datasets with minimal model size.
Camouflaged object detection (COD) aims to segment objects whose appearance is highly similar to the surrounding background. Its main challenges arise from low target-background discriminability and ambiguous object boundaries. Existing coarse-to-fine methods often depend on the quality of the initial coarse map, so localization errors may be propagated to subsequent prediction stages. To address this issue, we propose CMFR, a coarse map feedback refinement network with spatial-frequency feature integration. CMFR first extracts multi-scale contextual features for coarse target localization, then combines spatial and Laplacian frequency cues to aggregate boundary and texture responses, and finally uses coarse maps as dynamic thresholds for adaptive refinement. An RK2-inspired contextual integration module further stabilizes cross-scale feature refinement by combining immediate and corrected residual information through a learnable gate. Experiments on CAMO, CHAMELEON, COD10K, and NC4K show that CMFR achieves competitive performance compared with 19 recent COD methods, suggesting that coarse-map feedback and spatial-frequency refinement are useful for recovering ambiguous camouflaged regions. Our code is publicly available at: https://github.com/AaHa123/CMFR .
Accurate and real-time UAV-bird recognition is essential for anti-UAV systems, especially under resource-constrained edge deployment scenarios. In this paper, RT-DETR-MGC is proposed as a lightweight yet discriminative RT-DETR-based object detection framework for UAV-bird recognition. It emphasizes preserving discriminative feature representation while reducing computational redundancy. First, a MobileBlock-based backbone is constructed to decouple spatial feature extraction and channel interaction, while a C3Ghost-based feature fusion module enhances multi-scale feature interaction with minimal redundancy. In addition, a structured channel streamlining mechanism is designed to further reduce redundant channel transformations and improve model efficiency. Finally, extensive experiments on two public benchmarks and a self-constructed UAV-bird dataset demonstrate that compared with the baseline RT-DETR-R18, the proposed RT-DETR-MGC reduces the parameter count by 82.9
Precision livestock farming (PLF) plays a pivotal role in advancing animal welfare and health management, particularly in complex group-housed environments where monitoring the behavior of individual cows is challenging. Existing systems often struggle with scalability, detection accuracy, and the handling of occlusions in densely populated spaces. This paper introduces a behavior-aware monitoring framework that combines YOLOv8 for robust cow detection, EfficientNetV2-B1 for behavior classification, and an enhanced ByteTrack algorithm for multi-object tracking. The framework adopts a two-stage pipeline in which the first stage detects and isolates individual cows, and the second stage classifies their behavior into four categories: drinking, feeding, lying down, and standing. The proposed Cow-ByteTrack algorithm, an extension of ByteTrack, integrates behavior labels into the tracking pipeline to preserve both identity and behavioral continuity across video frames. Evaluated on a dataset of 1674 annotated images, the framework achieved a mean average precision (mAP@0.5) of 0.82 for detection, a behavior-classification accuracy of 94.12
Underwater video data routinely suffer from color distortion, low contrast, and frame-to-frame flicker due to wavelength-dependent absorption and scattering. Conventional single-frame UIE methods enhance individual images but introduce temporal artifacts when applied to video. We present T-PhISH-Net, a temporally consistent underwater enhancement model that augments the physics-aware PhISH-Net with a causal transformer-based temporal encoder, depth-guided feature fusion, and adaptive loss weighting. A monocular depth estimate (MiDaS DPT-Large) conditions the bilateral illumination grid via FiLM modulation and window-attention gating, improving backscatter–illumination separation in a physically motivated way. Extensive evaluations on Brackish, MVK, UOT32, and VDD-C show strong temporal stability (tSSIM, tPSNR) competitive with state-of-the-art methods, while maintaining competitive per-frame quality (UIQM). Ablations confirm the joint importance of causal attention, depth conditioning, and learned loss weighting. Limitations include dependence on monocular depth quality, added computational overhead from the MiDaS forward pass, and dataset-specific variation in which metric dimension leads. The method requires no optical flow and supports online inference, making it practical for AUV/ROV pipelines and long-term marine monitoring.
Most image inpainting methods operate on full images. In object-centric restoration, this design can introduce unnecessary background information into the reconstruction process. We study Specific Object-of-Interest Imaging (SIOI), a lightweight front-end module that produces object-level structural priors before downstream inpainting. The module is intended for cases in which the target object, rather than the surrounding scene, is the main subject of restoration. Experiments on a self-collected Teapot dataset and Common Objects in Context (COCO)-derived elephant, giraffe, and zebra subsets indicate that the proposed front end improves object-region reconstruction and boundary stability for several inpainting backbones, although the gains are not uniform across all metrics or categories. Ablation studies, robustness checks, loss-weight sensitivity, recent baseline comparisons, residual-error analysis, and a per-image audit of archived Teapot outputs are used to characterize where the module helps and where its limitations remain. These results support SIOI as a limited object-focused preprocessing component for restoration rather than a general full-image replacement.
Objective measures of evidence are critical in digital video forensics for establishing the authenticity and integrity of multimedia content. However, the development and benchmarking of such measures are hindered by the lack of comprehensive, real-world datasets that reflect the complexities of outdoor environments and splicing-based tampering. Consequently, video forensics approaches for detecting video tampering have received much attention in the forensics field in terms of crime scene investigations and are also presented as courtroom evidence. However, research on video-based splicing detection in outdoor scenes is lacking. In this paper, we introduce a new large-scale annotated spliced video dataset entitled “Outdoor Degraded Spliced Video Dataset (ODSVD)” that provides real-world outdoor-scenes under challenging atmospheric conditions. The proposed dataset contains spliced and corresponding authentic video clips of real-world scenes in atmospheric conditions, i.e., Fog, Rain, Clear Day, Clear Night, Night + Rain, Night + Fog, and Rain + Wind conditions. Along with these conditions, other challenging conditions, including acentric distribution of spliced objects, inter/intra class variability of spliced objects, and metadata tag-independent forged scenes, are also present in the dataset. The dataset contains 300 spliced and 300 corresponding original video clips in different atmospheric conditions. Additionally, 100 spliced video clips compressed using various social network sites (YouTube, Facebook, and Instagram) are included in the ODSVD dataset. In total, the dataset contains approximately 11.34 million frames (2700 to 4500 frames per video) in combination with the spliced and original clips. To make the dataset more efficient, the ground truth of spliced objects in the extracted frames is defined in the form of bounding boxes and binary masks. Using the ODSVD dataset, a quantitative comparison of the state-of-the-art splicing detection methods is performed. Unsurprisingly, the study reveals that the existing splicing detection techniques perform poorly when challenging atmospheric conditions are present, and there is still considerable scope for further improvement.
Medical images shared for collaborative diagnosis, remote consultation, and cloud-based analysis are exposed to the risk of unauthorized diagnostic inference, in which an unauthorized classifier extracts correct diagnostic predictions from a shared image without authorization from the data controller. This work addresses this specific threat through selective access control over classification outputs, and does not address anonymization, patient re-identification prevention, or broader privacy guarantees, which represent distinct threats outside the scope of the present framework. Existing adversarial example-based methods for restricting unauthorized inference suffer from three critical limitations: they fool all models indiscriminately including authorized users, they may corrupt diagnostically important lesion regions, and they are irreversible. To address these challenges, we propose a unified framework called Friend-Safe Semantic-Aware Reversible Adversarial Examples (FS-SA-RAE), which integrates three complementary mechanisms: (1) a friend-safe attack that employs dual-objective optimization to fool unauthorized (foe) models while preserving classification accuracy for authorized (friend) models, enabling selective access control; (2) semantic-aware perturbation that constrains adversarial modifications exclusively to non-lesion regions using segmentation masks, ensuring that diagnostically critical information remains intact; and (3) reversible data hiding based on histogram shifting that embeds deterministic perturbation-regeneration metadata into the protected image, enabling high-fidelity restoration of the original image when authorized. Comprehensive experiments on the ISIC2018 dermoscopy dataset ( n=1,503 test images), conducted with U-Net-predicted lesion masks rather than synthetic regions of interest, demonstrate that FS-SA-RAE attains a foe attack success rate above 74
Spodoptera frugiperda is one of the most economically important agricultural pests affecting maize, cotton, and soybean production worldwide. Efficient monitoring of egg masses is essential for population assessment, pest management decision-making, and biological control programs. This study proposes an automated framework for the detection and counting of S. frugiperda egg masses using a deep learning approach based on the YOLOv11 architecture. Six models were trained and evaluated under different configurations to investigate the effects of synthetic data augmentation, image resolution, loss-function weighting, and augmentation intensity on detection and counting performance. Model evaluation was performed using standard object detection metrics, including Precision, Recall, mAP@50, and mAP@50–95, as well as direct comparisons between automated predictions and manually annotated egg-mass counts. The best-performing configuration achieved a Precision of 88.90
Otolith classification plays an important role in fisheries science, supporting species identification, stock assessment, and marine resource management. However, the limited availability of labeled otolith images restricts the performance of conventional supervised deep learning models. In this study, we investigate three semi-supervised learning frameworks based on a Semi-Supervised Generative Adversarial Network (SGAN), a Semi-Supervised Standard Autoencoder (SSAE), and a Semi-Supervised Variational Autoencoder (SVAE) to improve otolith classification under limited labeled data conditions. To ensure a fair comparison, all methods are implemented using the same backbone architecture and are compared with a fully supervised CNN baseline. Beyond conventional evaluation metrics, we analyze the evolution of the decision boundary throughout training to assess class separability and model stability. Experiments conducted on the Moroccan Otolith Dataset demonstrate that the proposed semi-supervised approaches consistently achieve better performance than the supervised CNN. In particular, the SGAN, SSAE, and SVAE achieve classification accuracies of 94
