<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Artificial Intelligence | Pattern Recognition section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence/sections/pattern-recognition</link>
        <description>RSS Feed for Pattern Recognition section in the Frontiers in Artificial Intelligence journal | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-08-25T12:13:22.134+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1930978</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1930978</link>
        <title><![CDATA[Editorial: Deep learning for computer vision and measurement systems]]></title>
        <pubdate>2026-08-25T00:00:00Z</pubdate>
        <category>Editorial</category>
        <author>Seyed Jalaleddin Mousavirad</author><author>Mohammed El-Abd</author><author>Diego Oliva</author>
        <description></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1816534</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1816534</link>
        <title><![CDATA[A unified MAP–EM approach to stable Gaussian mixture clustering with priors, graphs, and split–merge adaptation for document clustering]]></title>
        <pubdate>2026-08-24T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sumathi Subbarayan</author><author>G. Hannah Grace</author>
        <description><![CDATA[IntroductionClustering high-dimensional and noisy data remains challenging for conventional expectation–maximization (EM) methods as overlapping clusters, sparse features, and outliers can lead to covariance degeneracy and unstable parameter estimates. This research aims to improve clustering performance in high-dimensional, noisy settings by developing a robust maximum a posteriori expectation–maximization (MAP–EM) framework that integrates prior regularization, geometric structure, and outlier handling. Traditional EM-based clustering methods often struggle in the presence of overlapping clusters, high-dimensional features, and outliers, leading to degenerate covariance and unstable parameter estimates.MethodsThe proposed MAP–EM model improves reliability by combining normal—inverse—Wishart (NIW) priors for covariance stabilization, a graph-Laplacian structure over the feature space to capture geometric relations among features, a uniform noise component to absorb outliers, and an adaptive split–merge strategy that refines cluster boundaries. These modules are coupled within a single MAP–EM procedure. The noise component modifies the E-step responsibilities by capturing atypical observations, and these updated responsibilities drive the NIW-regularized covariance and the graph-regularized mean updates in the M-step. The split–merge step is accepted only if it improves the penalized objective. The proposed MAP–EM model was evaluated on five synthetic datasets and three benchmark text corpora, namely Reuters-R8, BBC Sports, and the BBC dataset.ResultsOn Reuters-R8, the model achieved an Adjusted Rand Index of 0.347 and an accuracy of 0.569, outperforming the variational Bayesian Gaussian mixture model (GMM) (0.317). On the BBC dataset, it achieved the highest Adjusted Rand Index of 0.326 and a Normalized Mutual Information of 0.425 among the methods compared. Formal statistical testing showed that MAP–EM achieved significant positive differences in 45 out of 75 method-level comparisons, with one significant negative comparison. At the run level, MAP–EM obtained higher scores in 891 out of 1,115 valid paired comparisons, corresponding to a win rate of 79.9%. Theoretical analysis further supports the proposed framework by establishing coercivity of the penalized objective, monotone ascent of the MAP–EM iterations, finite termination of the split–merge stage under the stated acceptance criterion, boundedness of the covariance estimates under the normal–inverse–Wishart prior, local R-linear convergence of the MAP–EM iterations after model-order stabilization, eigenvalue bounds for the NIW covariance estimator, a condition-number bound for the penalized mean update, and a perturbation bound for the graph-regularized mean update.DiscussionThe proposed MAP–EM framework provides a stable, structure-aware clustering approach for high-dimensional text data. Experimental results indicate that its advantages are most evident on datasets with noise, overlapping clusters, and well-connected feature graphs, rather than across all clustering scenarios.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1838460</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1838460</link>
        <title><![CDATA[SE-ResAutoNet: an attention and residual learning framework for reliable gait-based suspect identification]]></title>
        <pubdate>2026-08-24T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Vaishnavi Munusamy</author><author>Sudha Senthilkumar</author>
        <description><![CDATA[Gait recognition relies on the distinctive walking pattern of the individuals that helps to identify the person accurately even in the long distance or in varied environments without any user interaction. Unlike existing techniques, the method works independently, it does not depend on a person engaging or interacting directly with the system. The SE-ResAutoNet (Squeeze-and-Excitation Residual Autoencoder Network) proposed in this study, an autoencoder framework which adopts SE blocks and Residual Blocks in their architecture for better feature representation and classification accuracy. The attention based - residual learning helps to capture the complex pattern effectively. The SE-ResAutoNet model has achieved an average accuracy of 95.75% across 5-fold cross-validation under normal, slow, and fast walking conditions, along with high precision, recall, and Rank-1 identification performance, demonstrating the consistency and robustness of the model. Furthermore, the proposed framework achieved benchmark accuracies of 97.1% on CASIA-B and 96.32% on OU-MVLP datasets, indicating strong cross-dataset generalization capability. The model also demonstrated computational efficiency with a minimal inference time of 11 ms per frame (77 FPS) and a compact computational cost of 1.5 GFLOPs, highlighting its suitability for real-world video-based surveillance applications. In addition, inclusion of Integrated Gradients highlights the most influential gait regions in the frame. The finding suggests that SE-ResAutoNet achieves an effective balance between accuracy, computational efficiency, and explainability, highlighting its strong potential for deployment in practical biometric and surveillance applications, subject to further validation on larger and more diverse datasets.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1769436</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1769436</link>
        <title><![CDATA[Basis-driven learnable operator for MLP-mixers]]></title>
        <pubdate>2026-08-24T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ahmed Elsheikh</author><author>Mohammed E. Fouda</author><author>Ahmed M. Eltawil</author>
        <description><![CDATA[IntroductionExisting multi-layer perceptron (MLP)-mixer architectures either rely heavily on extensive training data or employ rigid, hand-engineered mixing operations. This study introduces a Basis-Driven Learnable Operator (BDLO) for MLP-mixers, targeting a balance between high-capacity flexible learning and computationally efficient, yet constrained, handcrafted solutions.MethodsBDLO is a drop-in replacement for the shifting block of shifting-based MLP-mixers. It approximates shifting-based mixing operations by learning the coefficients of a real, complete, discrete basis, yielding an input-dependent transformation matrix that is applied to both rows and columns of the token table, while the patch embedding, channel-mixing MLPs, skip connections and classification head are left unchanged. The operator was integrated into CycleMLP, HireMLP and AS-MLP at three model sizes each, and all models were trained from scratch under an identical configuration on CIFAR10, CIFAR100, a reduced ImageNet (32 × 32, 500 classes) and full ImageNet1K. Standard and discrete cosine transform bases were compared, hyperparameter sensitivity was assessed with Optuna, and differences were tested using the Wilcoxon signed-rank test.ResultsBDLO reduced the cost of the mixing layer to 1.68 GFLOPs, against 3.33–5.08 GFLOPs for the original layers, and reduced whole-model parameter counts by 21.8%–56.7% (mean reduction 12.84M), with a mean accuracy difference of 0.38% in favor of BDLO. The Wilcoxon test confirmed a significantly lower parameter distribution (p = 0.0039) and no statistically significant accuracy difference (p = 0.496 on CIFAR100, p = 0.0625 on CIFAR10, p = 0.5 on the reduced ImageNet). Results were invariant to the choice of basis (mean cosine distance 0.0049 between the learned coefficient vectors) and insensitive to the BDLO-specific hyperparameters.DiscussionComparability holds in aggregate and in the parameter-constrained regime, with model-specific exceptions for baselines that include channel mixing (HireMLP) and for high-capacity baselines on larger datasets (AS-MLP). BDLO behaves as an input-adaptive spectral modulator whose bounded coefficients provide implicit regularization. This confirms the effectiveness of BDLO as an efficient operator replacement for shifting-based MLP-mixers, most suitable for parameter-constrained, small-to-medium-scale models in image recognition tasks.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1872978</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1872978</link>
        <title><![CDATA[Quantum-enhanced deep learning models for automated defect detection in materials]]></title>
        <pubdate>2026-08-20T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aniket Garg</author><author>Snehal Sushibine</author><author>Diya Choudhuri</author><author>Shrey Anand</author><author>Helen Vijitha P.</author><author>Jai Vinita L.</author><author>Parvathy A. K.</author><author>Vinaytosh Mishra</author><author>Thompson Stephan</author>
        <description><![CDATA[The quality of raw materials is fundamental to the reliability and overall performance of final products, serving as the cornerstone of modern manufacturing standards. While traditional inspection methods can be effective, they are frequently time-consuming, labor-intensive, and unsuitable for high-throughput production environments. Recent advances in quantum computing offer significant advantages for processing large-scale data and complex feature representations, with strong potential to address the limitations of traditional inspection techniques. This study introduces an automated approach for grading raw materials, such as steel and fabric, which are widely used in industries including automotive, textiles, and general manufacturing, by integrating deep learning with quantum machine learning. A Hybrid Quantum Classical Neural Network (HQCNN) is proposed to enhance defect detection. The HQCNN utilizes a fine-tuned ResNet-18 backbone for robust feature extraction. Extracted features are used by a trainable depth controller that determines the complexity of a six-qubit variational quantum circuit (VQC) with data-adaptive, dynamic depth. Each VQC layer consists of rotations around the Y-axis (RY) and Z-axis (RZ), combined with Controlled-NOT (CNOT) gates arranged in a ring topology, with the total number of layers determined on a per-sample basis. The model is trained using a resource-aware loss function that penalizes excessive circuit depth, and the resulting quantum features are fused with classical features for final classification. Comprehensive ablation studies comparing the HQCNN against parameter-matched classical multi-layer perceptrons (MLPs) and fixed-depth quantum circuits demonstrate the fundamental advantage of the proposed approach. The dynamic HQCNN achieves 88.1 ± 2.4% accuracy on the complex metal surface defect dataset and 93.0 ± 1.8% on the fabric defects dataset. More importantly, it yields a statistically significant improvement in the macro F1-score (from 0.79 in classical parameter-matched baselines to 0.87), proving its superior capability in recognizing complex, minority-class defects while utilizing a fraction of the trainable parameters in the classification head compared to traditional architectures.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1800342</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1800342</link>
        <title><![CDATA[Evaluating the robustness of specialized and general-purpose facial expression recognition systems across varied scenarios]]></title>
        <pubdate>2026-08-17T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>José Salas-Cáceres</author><author>Javier Lorenzo-Navarro</author><author>Modesto Castrillón-Santana</author><author>Patricia Picazo-Peral</author><author>Sergio Moreno-Gil</author>
        <description><![CDATA[IntroductionThis work presents a comprehensive evaluation of facial expression recognition (FER) systems across four benchmarked datasets of varying complexity, ranging from controlled static images (ADFES, WSEFEP) to more realistic dynamic recordings (RAVDESS, CREMA-D).MethodsThree categories of models were evaluated: traditional FER neural networks models, general-purpose vision language models (VLMs), and the commercial software FaceReader© 10.ResultsThe results show that performance on controlled datasets substantially overestimates real-world FER capability, with average weighted and unweighted average recall values decreasing from approximately 72% in static datasets to below 30% in naturalistic settings. All tested models exhibited a marked bias toward happiness, with negative emotions frequently misclassified, a trend particularly pronounced in VLMs, where categories such as fear or anger often received F1-scores near zero. Among the neural networks, the DAN model trained on the AfectNet dataset achieved the strongest generalization, outperforming all VLMs and confirming that AfectNet provides a more realistic training distribution than the RAF-DB database. FaceReader© delivered excellent performance under ideal conditions but experienced substantial degradation in dynamic scenarios, falling below a random classifier in CREMA-D.DiscussionThese findings highlight the limitations of general-purpose VLMs and commercial tools for real-world FER and underscore the need for models explicitly designed to handle naturalistic variability. Furthermore, the reported performance of FaceReader© 10 in their manual on ADFES and WSEFEP was corroborated in this study.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1817529</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1817529</link>
        <title><![CDATA[Reassessing demographic bias in face attribute classification: a statistically grounded multi-model evaluation on FairFace and UTKFace]]></title>
        <pubdate>2026-08-14T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Andisani Nemavhola</author><author>Serestina Viriri</author><author>Colin Chibaya</author>
        <description><![CDATA[Face analysis systems are widely used in security, authentication, and public-sector applications; however, demographic bias and the statistical reliability of reported performance remain key concerns. Many studies rely on aggregate accuracy without quantifying subgroup disparities or uncertainty, potentially overstating model fairness. This study presents a statistically grounded evaluation of demographic bias in face attribute classification across three representative architectures, ResNet50, MobileNetV3, and a vision transformer (DeiT), using the FairFace and UTKFace datasets. Subgroup analysis is conducted across race and gender, incorporating disparity indices, bootstrap confidence intervals, and inferential statistical testing with effect size analysis. The evaluation uses an embedding-based nearest-neighbor approach to examine representation-level behavior consistently across models. Results show that race-based disparities are substantially larger than gender-based disparities across both datasets. On FairFace, race disparity gaps range from 0.1124 to 0.1266, while on UTKFace they increase significantly to 0.4726–0.4944, with large effect sizes (Cohen's d>1). In contrast, gender disparities remain smaller, with gaps between 0.0280 and 0.0582 on FairFace and 0.0194–0.0326 on UTKFace, and correspondingly small effect sizes (d < 0.13). Despite modest differences in overall accuracy across models, subgroup disparities remain statistically significant across all architectures. These findings emphasize the importance of subgroup-level evaluation, uncertainty quantification, and statistical validation for reliable fairness assessment in face analysis systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1897208</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1897208</link>
        <title><![CDATA[Emotion recognition from body movement through interpretable motion-aware sequential modeling]]></title>
        <pubdate>2026-08-06T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sergio Esteban-Romero</author><author>Iván Martín-Fernández</author><author>Rubén San-Segundo</author><author>Manuel Gil-Martín</author><author>Fernando Fernández-Martínez</author>
        <description><![CDATA[Emotion recognition from bodily movement remains a challenging problem, particularly when only pose-based motion sequences are available and emotionally informative content is not uniformly distributed across time. In this work, we propose a Window Transformer architecture grounded in the Multiple Instance Learning (MIL) paradigm to address this challenge. Rather than processing the full sequence as a single temporal stream, the model decomposes it into overlapping windows and learns to assign greater relevance to those segments containing stronger emotional content. This formulation provides a more interpretable framework, since the learned relevance scores reveal which temporal regions drive the final prediction, while also yielding richer, context-aware representations of each segment. We evaluate the proposed approach on two publicly available datasets, MEED and DIEM-A, and compare it against a standalone Transformer baseline under different batch size and window configuration settings. The Window Transformer consistently outperforms the baseline and exhibits a more stable behavior across training configurations, achieving best accuracies of 56.35 ± 2.67% on MEED and 23.84 ± 0.83% on DIEM-A in a subject-independent scenario. Beyond performance gains, this work also establishes new benchmark values on both datasets, providing reference results that future research on bodily emotion recognition can build on.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1816796</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1816796</link>
        <title><![CDATA[Hybrid CNN with angular margin supervision for robust face identification and verification]]></title>
        <pubdate>2026-07-27T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Andisani Nemavhola</author><author>Colin Chibaya</author><author>Serestina Viriri</author>
        <description><![CDATA[Face recognition systems are widely used in surveillance, biometric authentication, access control, and digital identity verification; however, supervision sensitivity, evaluation stability, and performance consistency across datasets remain insufficiently understood. This study investigates the behavior of convolutional, transformer-based, and hybrid face recognition architectures under both Softmax and ArcFace supervision using five-fold subject-disjoint cross-validation on the Labeled Faces in the Wild (LFW) and FAGEv2 datasets. ResNet50, MobileNetV3, DeiT-Small, and a Hybrid multi-branch architecture integrating complementary convolutional and transformer feature representations were evaluated using Top-1 identification accuracy, Area Under the ROC Curve (AUC), Equal Error Rate (EER), computational complexity, and fold-level statistical analysis. Experimental results revealed substantial supervision sensitivity across architectures and datasets. On the LFW dataset, Hybrid-Softmax achieved the highest Top-1 identification accuracy (62.4%), while DeiT-Small-Softmax achieved the strongest verification performance with an AUC of 0.905 and EER of 0.159. On the FAGEv2 dataset, Hybrid-Softmax and DeiT-Small-Softmax achieved the highest identification accuracy (38.0%), while Hybrid-Softmax achieved the strongest verification performance with an AUC of 0.825 and EER of 0.251. Fold-level analyses demonstrated that the effect of ArcFace supervision varied across architectures and datasets, with consistent improvements observed for some convolutional architectures but not for transformer-based or hybrid models. Cross-dataset evaluation further revealed changes in model ranking and supervision behavior, indicating that comparative performance is strongly influenced by dataset characteristics and evaluation conditions. The findings demonstrate that additive angular margin supervision does not universally outperform conventional Softmax optimization and highlight the importance of multi-dataset benchmarking, fold-level evaluation, and supervision sensitivity analysis for robust and reproducible face recognition benchmarking.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1858540</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1858540</link>
        <title><![CDATA[PSO-based parameter optimization of intuitionistic fuzzy generator for low-light image enhancement]]></title>
        <pubdate>2026-07-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Uma Maheswari S.</author><author>Jagatheswari S.</author>
        <description><![CDATA[Low-light images often suffer from reduced visibility, noise, and loss of structural details due to insufficient illumination and sensor limitations. These degradations affect both visual perception and downstream image analysis tasks. This paper presents a low-light image enhancement framework based on intuitionistic fuzzy generator (IFG) integrated with gamma correction and optimized using particle swarm optimization (PSO). As a preprocessing step, block-matching and 3D filtering (BM3D) are applied to suppress noise while preserving structural information. The IFG models uncertainty in pixel intensities to enable adaptive contrast enhancement, whereas gamma correction adjusts brightness levels. The enhancement parameters are optimized using PSO guided by dataset-specific objective functions, namely structural similarity (SSIM) for reference datasets and entropy-based optimization for no-reference scenarios where ground-truth images are unavailable. Experimental evaluations on standard benchmark datasets using both reference and no-reference image quality metrics indicate that the proposed framework achieves competitive enhancement performance with improved contrast and preservation of visually relevant image details. Although the computational cost is higher than that of feed-forward deep learning models, the framework is suitable for applications where training data are unavailable and interpretable parameter-adaptive enhancement is preferred.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834376</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834376</link>
        <title><![CDATA[Towards synthetic fillers for fair photo lineups: application of generative AI in criminal law proceedings]]></title>
        <pubdate>2026-07-07T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>A. Dörsch</author><author>R. Nichols</author><author>C. Busch</author><author>C. Rathgeb</author>
        <description><![CDATA[A photo lineup is an identification procedure widely used in criminal investigations. It involves presenting a suspect embedded into a set of known-innocent candidates (fillers) to an eyewitness in order to identify the suspect. Compiling fair lineups remains a challenge, particularly when the investigation team has to resort to unsuitable fillers. To support real-world lineups and mitigate the risk of misidentification, we present a practical approach for generating synthetic fillers. By injecting weighted layer-wise noise into a suspects latent vector representation, our approach generates visually distinct fillers while largely preserving demographic characteristics of the suspect. To assess suitability for investigative scenarios, we conducted a large-scale human perception study involving over 450 participants. The results show that the use of synthetically generated fillers leads to balanced identification performance, without making the suspect indistinguishable or stand out. Further experiments show that, compared to previous work, our approach achieves improved preservation of demographic characteristics. Overall, our work contributes to improving fairness in lineups and opens up avenues for supporting criminal investigations through the use of synthetic data.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1878911</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1878911</link>
        <title><![CDATA[Symmetry-constrained hybrid quantum-classical convolutional neural networks for rotation-robust face recognition]]></title>
        <pubdate>2026-07-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>S. Sony Priya</author><author>R. I. Minu</author>
        <description><![CDATA[Face recognition systems struggle when faces appear at different orientations. Standard convolutional neural networks handle translation well but have no built-in way to deal with rotations or reflections. Quantum neural networks offer a different kind of expressiveness, but most existing designs ignore spatial symmetry altogether. This paper introduces Eq-MG-QCNN, a hybrid quantum–classical model that builds rotation symmetry directly into the quantum circuit. The quantum filter is designed to be exactly equivariant under the Klein four-group, which covers horizontal flips, vertical flips, and 180° rotations. This is done through two mechanisms: sharing rotation parameters across all qubits (as required by orbit analysis), and connecting all qubit pairs with symmetric CZ gates (forming a complete K₄ graph). The model is tested on the ORL and Yale face databases under four rotation angles (0°, 90°, 180°, 270°) and compared against a classical CNN, a classical equivariant CNN, and the MG-QCNN quantum baseline. All experiments use noiseless quantum simulation. Eq-MG-QCNN reaches 94.3% best accuracy on ORL and 89.9% on Yale with only six quantum parameters, two fewer than the baseline. The model also shows low rotational variation across all four test angles. These results suggest that embedding group symmetry into quantum circuits is a practical way to build orientation-stable feature extractors for face recognition.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1804734</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1804734</link>
        <title><![CDATA[Compact waste image classification with multi-student CNNs and edge-oriented model selection]]></title>
        <pubdate>2026-06-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mohamed Echchidmi</author><author>Anas Bouayad</author>
        <description><![CDATA[Automatic waste classification is an important enabling technology for cleaner cities, source-level recycling, and low-cost smart-bin systems. Although modern convolutional neural networks achieve strong recognition performance, their deployment on affordable edge devices remains constrained by memory footprint, computational cost, and response latency. This paper presents an edge-oriented compact CNN framework for waste image classification, combining a high-accuracy MobileNetV4 reference model with three lightweight student architectures: EfficientNet-Lite0, LCNet-0.5, and MobileNetV3-Small-0.5. All models are evaluated on TrashNet under a unified preprocessing, training, and size-accounting protocol, allowing a clear comparison of accuracy–efficiency trade-offs. On the main stratified train/validation/test split, the MobileNetV4 teacher achieves 97.09% top-1 accuracy, while the compact students retain strong performance with substantially smaller footprints: EfficientNet-Lite0 reaches 93.99% with 3.38 M parameters, LCNet-0.5 reaches 94.18% with only 0.61 M parameters, and MobileNetV3-Small-0.5 reaches 87.73% with 0.57 M parameters. A complementary stratified five-fold evaluation, including both knowledge-distilled and non-distilled student variants, provides a robust assessment of model behavior across data partitions and confirms LCNet-0.5 as the most suitable sub-megabyte candidate under the proposed size–accuracy selection rule. The selected LCNet-0.5 model achieves a macro-F1 score of 0.9247 on the main TrashNet test split and is integrated into a self-contained Raspberry Pi 3 Model B+ prototype that performs local camera-to-display inference with an observed end-to-end latency of approximately 1.0 s per image. Cross-dataset evaluation on RealWaste further shows that the compact model can be adapted effectively to cluttered real-world imagery through short fine-tuning. Overall, the results demonstrate that careful lightweight architecture selection, supported by knowledge distillation analysis and edge-prototype validation, can deliver accurate, compact, and practically deployable waste classifiers for resource-constrained environments.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1835651</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1835651</link>
        <title><![CDATA[Evaluating the real-world robustness of face-swap detection models under compression and noise]]></title>
        <pubdate>2026-06-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Li Baocai</author><author>Fazli Bin Azzali</author><author>Nik Fatinah binti N. Mohd Farid</author>
        <description><![CDATA[IntroductionRecent advances in generative adversarial networks (GANs) and autoencoding techniques have significantly improved the realism of face-swap and deepfake media, creating substantial challenges for digital media authentication. Although existing deepfake detection models achieve high accuracy on benchmark datasets, their robustness under real-world media degradations remains insufficiently explored.MethodsThis study systematically evaluates the resilience of five leading face-swap detection models—XceptionNet, MesoNet, FSD-GAN, FakeTracer, and a Hybrid + Landmark approach—under four common distortions: JPEG compression (quality levels 20–90), Gaussian noise (σ = 0.01–0.05), motion blur (kernel size 3–15), and video encoding artefacts (bitrate 50–500 kbps). Experiments were conducted using the FaceForensics++ dataset (1,000 videos: 720 training, 140 validation, and 140 testing) and Celeb-DF v2 (590 videos: 400 training, 90 validation, and 100 testing). Performance was assessed using accuracy, F1-score, area under the curve (AUC), and degradation rate (Δ) between clean and distorted conditions.ResultsThe results demonstrate a substantial reduction in detection performance under degraded conditions. Average accuracy declined from 94.7% on clean data to 67.8% on distorted data, corresponding to an overall degradation rate of −26.9%. JPEG compression and motion blur caused the most significant performance drops, with reductions of up to 35%, particularly for lightweight CNN-based detectors. In contrast, FSD-GAN and FakeTracer exhibited greater robustness, maintaining degradation rates of no more than −15% due to their latent fingerprinting and trace embedding mechanisms.DiscussionThe findings highlight the limitations of current deepfake detection systems when deployed in real-world environments where media distortions are prevalent. The study emphasizes the need for distortion-aware training strategies, cross-condition benchmarking, and deployment-oriented evaluation protocols. Furthermore, a dual-branch framework integrating a Vision Transformer (ViT) for spatial artefact detection with a Recurrent Neural Network (RNN) or Temporal Convolutional Network (TCN) for temporal coherence modelling is proposed as a promising direction for improving the robustness and reliability of future deepfake detection systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1833234</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1833234</link>
        <title><![CDATA[Violation detection in power operation sites based on multi-scale detection and few-shot learning]]></title>
        <pubdate>2026-06-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yaokuan Wen</author><author>Jun Wang</author><author>Qiming Liu</author><author>Mo Zhou</author><author>Minzhe Tian</author>
        <description><![CDATA[IntroductionSafety supervision at power operation sites is critical for ensuring worker safety and maintaining a reliable electricity supply. However, existing safety violation detection methods are constrained by limited labeled data, poor performance on small-object detection tasks, and interference from complex backgrounds.MethodsTo overcome these challenges, this study proposes a framework that integrates multi-scale object detection with few-shot learning. A multi-scale feature extraction module is designed based on a feature pyramid network and channel attention mechanisms to enhance the perception of small objects. In addition, a few-shot learning framework incorporating a meta-learning strategy is introduced to address the scarcity of labeled safety violation samples and improve the model's adaptability to new tasks with limited training data.ResultsExperimental results demonstrate that the proposed method consistently outperforms existing approaches across multiple evaluation metrics. The framework achieves notable improvements in small-object detection accuracy and few-shot learning performance, resulting in enhanced detection accuracy, robustness, and generalization capability.DiscussionThe integration of multi-scale feature extraction and few-shot learning effectively addresses the challenges of safety violation detection in power operation environments. The proposed framework provides a practical and reliable solution for intelligent safety monitoring and has significant potential for real-world deployment in power operation sites.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1834763</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1834763</link>
        <title><![CDATA[Adaptive quadtree-based segmentation of nucleus and cytoplasm in pap-smear images: a lightweight and interpretable approach for automated cytology]]></title>
        <pubdate>2026-06-12T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Wasswa William</author><author>Andrew Ware</author>
        <description><![CDATA[BackgroundAutomated analysis of Pap-smear images plays an important role in cervical cancer screening, particularly in low-resource settings where manual cytology remains labour-intensive, subjective, and prone to inter-observer variability. On the other hand, accurate segmentation of the nucleus and cytoplasm is a fundamental step in computer-aided diagnosis systems because it enables quantitative morphometric analysis and computation of clinically important biomarkers such as the nucleus-to-cytoplasm ratio. However, robust cervical cell segmentation remains challenging due to staining variability, inhomogeneity, irregular morphology, weak cytoplasmic boundaries, and overlapping cellular structures. This study presents an adaptive quadtree-based segmentation framework for automated nucleus and cytoplasm delineation in Pap-smear images.MethodsThe proposed method employs hierarchical split–merge decomposition guided by a dynamic adaptive statistical homogeneity analysis using mean intensity, variance, and entropy measures. Preprocessing is performed using large-kernel median filtering for background normalisation, followed by local Otsu thresholding, adaptive region merging, overlap refinement, and morphological post-processing. The framework was evaluated on both the Herlev cervical cytology dataset and the ISBI 2015 cervical cytology segmentation challenge dataset containing overlapping and clustered cervical cells. Comparative benchmarking was additionally performed against the U-Net and Attention U-Net.ResultsOn the Herlev dataset, the proposed framework achieved nucleus Dice coefficients exceeding 0.94 and Zijdenbos Similarity Index (ZSI) values greater than 0.9034 across all diagnostic classes, with competitive cytoplasm segmentation performance. On the ISBI 2015 dataset, the framework maintained acceptable segmentation performance under overlapping-cell conditions, achieving nucleus Dice and ZSI values of 0.912 ± 0.048 and 0.918 ± 0.044, respectively. Morphometric feature comparisons demonstrated strong agreement with ground-truth annotations and low average percentage errors for area and diameter measurements. Although deep learning models achieved superior performance under highly complex overlap conditions, the proposed framework remained competitive while requiring substantially lower computational resources and no iterative model training.ConclusionThe proposed Adaptive Quadtree-Based Segmentation framework provides a lightweight, interpretable, and computationally efficient approach for automated cervical cytology segmentation. Its training-free design, transparent statistical decision rules, and reduced hardware requirements make it particularly suitable for deployment in resource-constrained and embedded cervical cancer screening systems. The framework provides a practical segmentation backbone for automated cytology analysis and downstream computer-aided diagnosis applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1841848</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1841848</link>
        <title><![CDATA[An efficient hybrid CNN–transformer framework for real-time weapon detection and face recognition]]></title>
        <pubdate>2026-06-10T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>P. Shanthi</author><author>V. Manjula</author>
        <description><![CDATA[The growing demand of smart surveillance systems necessitates the accurate and real-time detection of weapons and face recognition with robustness against occlusion, illumination changes, and complex backgrounds. Existing techniques based on standalone CNN or transformer architectures are less effective in capturing local fine-grained features as well as long-range dependencies. This paper presents ConViDeTR, a hybrid deep learning framework that integrates CNN, Vision Transformer (ViT), and Detection Transformer (DETR) architectures into a unified framework. The key contribution of the proposed framework is the deep feature fusion layer, which integrates local spatial features, global context features, and object query features in one shared feature space. This enables the synchronous execution of weapon detection and face recognition using one efficient framework. The experiments using existing benchmark datasets validate the performance, achieving 98.9% accuracy in weapon detection and 97.34% accuracy in face recognition and outperforming the existing techniques in both tasks. The framework also demonstrates real-time performance with 25–30 FPS and low latency. The performance of the proposed framework sustains its effectiveness, robustness, and scalability in the development of next-generation intelligent surveillance systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1784359</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1784359</link>
        <title><![CDATA[Deep learning and multi-statistical features: an intra-frame forgery detection video method]]></title>
        <pubdate>2026-05-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Diaa Uliyan</author><author>Manal Eid Alazmi</author><author>Mohammad Alsaffar</author><author>Moham’d Al-Dlalah</author><author>Meshari Alazmi</author>
        <description><![CDATA[IntroductionIn this paper, we propose a technique for detecting spliced video forgeries using statistical clues extracted from the spatial and compression domains inside a suspicious video. The proposed technique employs a multi-feature architecture to train statistical features for every domain using the VGG-16 model. Examining the fused characteristics in specific image regions with great detail exposes instances of manipulation. Both the compression impacts of JPEG and the visual distortions that result from image modification are examined in our research.TechniquesIn this study, we propose an alternative to the standard methods for detecting spliced forgeries in spatial domain textural analysis, such as entropy-based edges (MER), median filter residual (MFR), gray level regional maxima (FGM), morphological open images (MOI), and morphological erosion images (MEI). Derivative filters of image that may extract statistical information about manipulation traces from compressed data were chosen for artifact detection in double JPEG compression. As an example, the compression domain analysis makes good use of the 2D block Discrete Cosine Transform (DCT). There are three stages that comprise up the suggested method: (1) Video preprocessing splits videos into frames and converts each frame’s image to two domain formats: spatial and DCT compression. (2) Take each domain and extract several statistical characteristics. (3) Train VGG16 on a dataset to identify the spatially manipulated area in a video.ResultsThe proposed method has been validated and tested on the HTVD and GRIP datasets. The performance measure, such as Splicing forgery accuracy is 93.50% on the GRIP dataset and 92.40% in HTVD.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1824634</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1824634</link>
        <title><![CDATA[Redefining lightweight vision models for healthcare AI]]></title>
        <pubdate>2026-05-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Linus Lee</author><author>Zhibin Feng</author><author>Jen Hong Tan</author><author>Chiaw-Ling Chng</author>
        <description><![CDATA[BackgroundVision models for medical imaging often require tens of millions of parameters, raising questions about whether architectural efficiency can be achieved without sacrificing classification performance. We introduce MedLiT-seed (2.1 Million parameters) and MedLiT-nano (0.75 Million parameters), two ultra-lightweight vision transformers designed for efficient and scalable medical image analysis.MethodsMedLiT employs a streamlined Mixture-of-Experts (MoE) architecture with SwiGLU feedforward networks, grouped query attention, and depth-wise scaling. Models were pre-trained using masked autoencoding on ImageNet and MedMNIST, followed by fine-tuning on 12 MedMNIST 2D subsets. We evaluated performance across multiple configurations and compared against benchmark models including ResNet, MedViT, and AutoML systems.ResultsMedLiT-seed achieved the highest Area Under Curve (AUC) on 4 subsets and second-highest on 2 others, outperforming models with 10–20× more parameters. MedLiT-nano achieved results comparable to, and even exceeding, ResNet-18 and AutoML baselines in several subsets. Transfer learning from ImageNet significantly improved convergence and generalization. Increasing embedding size yielded greater performance gains than increasing expert count.ConclusionMedLiT demonstrates that MoE-based token routing represents a viable architectural pathway for achieving competitive accuracy relative to its floating-point operations (FLOP) across diverse medical imaging modalities on the order of 2M parameters. These results suggest that selectively routing computation through specialised experts, rather than scaling model size, can serve as an effective design principle for more compact medical vision models. Such architecture can be utilised for low-resource clinical environments and scalable fine-tuning across diverse healthcare tasks, though limitations on multi-label tasks highlight clear directions for future architectural refinement.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1809586</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1809586</link>
        <title><![CDATA[HybridWeaveNet: deep cultural pattern recognition for Indian handloom heritage fabrics]]></title>
        <pubdate>2026-05-28T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>S. A. Anuraag Shankar</author><author>A. Sasithradevi</author><author>G. Krishnaraj</author><author>S. Kanimozhi</author>
        <description><![CDATA[IntroductionIndia’s handloom sector represents the diversity of local culture, which is reflected in the traditional handwoven products from India. Regional fabric varieties such as Bandhani, Banarasi, Kancheepuram, Patola and Tussar are characterized by unique pattern structures, in terms of yarns as well as color arrangements. Accurate identification between them is crucial for tasks such as digital archiving, automatic categorization of fabric and e-commerce. Manually categorizing these designs would be tedious and subjective.MethodsFurther, the existing conventional deep learning and computer vision approaches encounter significant challenges when processing complex motifs, fine-grained textures and substantial intra-class variability. For this purpose, we propose HybridWeaveNet, a custom deep learning architecture specifically developed to mitigate these shortcomings. The model integrates both a Dual Attention mechanism and pretrained EfficientNetV2 backbone to enhance pattern recognition and variate feature learning. Training of the proposed model is carried out using balanced sampling among the available five-class in the dataset. Further, various augmentation methods, including Mixup, Cutmix, GridDropout, Cutout, Blur, and ColorJitter, have been employed for effective generalization.ResultsThe proposed HybridWeaveNet model is evaluated using various metrics such as accuracy, macro F1-score, Cohen’s kappa, Matthews correlation coefficient (MCC), Jaccard index and log loss and exhibited 91% performance.DiscussionGrad-CAM visualizations validated that the model concentrates on culturally related motifs, thereby highlighting its potential for extensive fabric classification.]]></description>
      </item>
      </channel>
    </rss>