This study presents an intelligent framework for identifying security intrusions around wind farms by integrating advanced video surveillance and target tracking technologies. To address the challenges posed by dynamic outdoor environments - such as occlusions, illumination changes, and large scale spatial layouts - the framework introduces two core components: The Dynamic Surveillance Intrusion Detection Model (DSIDM) and the Dynamic Intrusion Detection Framework (DIDF). The DSIDM leverages a hierarchical design combining convolutional neural networks for spatial feature extraction and recurrent structures for modeling temporal dependencies. It incorporates feature encoding, object localization, motion pattern understanding, and multiobject tracking to accurately recognize diverse categories, including authorized personnel, wildlife, and potential intruders. A decision mechanism further integrates spatial temporal cues to evaluate intrusion likelihood with high robustness. Complementing this, the DIDF employs adaptive surveillance principles to refine detection under real-world operational variability. It integrates spatial-temporal prediction, probabilistic modeling, reinforcement learning-based decision optimization, and noise reduction mechanisms to ensure reliable performance under fluctuating weather, camera noise, and complex motion behaviors. By dynamically adjusting thresholds and surveillance actions, the DIDF enhances responsiveness and reduces false alarms. Extensive experiments conducted on multiple wind farm-related surveillance datasets demonstrate the effectiveness of the proposed framework, showing improvements in accuracy, robustness, and computational efficiency compared with state of the art methods. Together, the DSIDM and DIDF provide a scalable and adaptive solution for real time protection of wind farm infrastructure.
Identifying brain stroke lesions from medical images is essential for timely diagnosis, treatment planning, and patient tracking. Deep learning-based methods have demonstrated significant opportunities for brain stroke detection in both Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). The location of the lesioned tissue can be obtained using automated stroke lesion segmentation, which can aid in clinical practice by assessing and evaluating the risks connected with different treatments. Hence, this review aims to discuss and clarify the present research work on sub-acute brain ischemia detection using CT and MR images. Multiple CT and MRI datasets utilized for the observation are gathered and included. Different experimental tools used for the validations are detailed. A chronological review related to the suggested models according to the contributions is provided. Various machine learning as well as deep learning mechanisms utilized for the validations are tabulated. Several advancements and complications associated with the existing research works are tabulated and included in this work. Multiple performance metrics utilized in the validations are tabulated. At last, various difficulties are presented in the classical technique that need to be tackled in the upcoming research and explained as the research gaps. Hence, the developed research work effectively helps the upcoming professionals gain more information related to their research.
Precise segmentation of pests in agricultural images is highly necessary for early infestation detection, effective pest monitoring, and significant crop protection. Anyhow, the variations in background complexity, pest appearance, and noise frequently minimize the precision and stability of traditional segmentation approaches. To overcome this challenge, this research proposes a deep learning-based pest segmentation framework that combines Adaptive Weighted Mean Filtering (AWMF) with the Global-Local Representation Learning Network (GlLo-SegNet). Initially, AWMF is applied using the proposed method to minimize the undesired biases and enhance image quality during preprocessing. Then, the GlLo-SegNet-based segmentation process is applied by integrating global semantic information and local geometric information using multi-scale convolution (MC), multi-scale pooling (MP), and multi-scale feature fusion strategies. Additionally, in order to enhance multi-scale fusion and semantic segmentation accuracy, an attention-based decoder is designed. The significance of the proposed approach lies in the capability to attain precise and stable pest segmentation in the case of challenging agricultural imaging conditions. The experimental outcomes indicate the model’s effectiveness on two benchmark datasets. On the agricultural pest dataset, the proposed method attained 97.76% recall, 98.74% precision, and 98.25% Dice coefficient. Using the crop pest and disease detection dataset, 97.26% recall, 98.82% precision, and 98.13% Dice coefficient are obtained. These outcomes ensure that the proposed model presents an effective and robust solution for agricultural pest segmentation.
Deep learning (DL) has emerged as a powerful tool for automated analysis of spine and intervertebral disc (IVD) imaging, offering promising solutions for detection, segmentation and classification tasks in magnetic resonance imaging and computed tomography modalities. Accurate interpretation of spinal images is clinically important but remains a challenging task due to anatomical complexity, image variability and inter-observer differences in manual assessment. In order to understand the applicability of DL models in spine and IVD imaging, this review provides a structured and critical overview of recent DL-based research articles. Unlike existing surveys, this work emphasizes a comparative analysis of model architectures, dataset sources and characteristics, evaluation metrics, validation strategies, and clinical applicability. Key challenges related to dataset heterogeneity, limited external validation, inconsistent performance reporting, and barriers to clinical adoption are systematically discussed. The review also highlights emerging research directions, including standardized benchmarks, multi-contrast and multimodal learning, explainable artificial intelligence, and robust cross-institutional validation to support the development of reliable and clinically deployable spine imaging systems.
Illegal image identification within network content review systems requires robust solutions capable of handling both complex multi-label classification and precise localization of sensitive regions. This study presents an integrated framework designed to enhance the accuracy and interpretability of automated moderation by combining a multi-label classification model with advanced spatial analysis. Central to this framework is the Sensitive Area Detection Network (SADN), which leverages convolutional feature extraction and multi-head attention to capture diverse semantic cues while generating refined bounding box predictions for regions requiring human-level scrutiny. SADN adopts a dual output structure that jointly optimizes label prediction and spatial localization, enabling comprehensive understanding of content containing multiple forms of illegal material. Complementing SADN is the Adaptive Content Review Strategy (ACRS), a dynamic mechanism that incorporates multimodal information, adaptive learning, and a scalable verification pipeline. ACRS integrates visual and textual cues through a fusion-based encoder, applies attention-guided refinement to highlight high risk regions, and adopts reinforcement-driven updates to improve decision making across evolving content types. Experimental evaluations conducted on multiple datasets demonstrate that the proposed framework significantly improves classification accuracy, reduces false positives, and enhances localization precision compared with state-of-the-art approaches. Together, SADN and ACRS provide a unified and scalable foundation for next generation content review systems, offering both high detection performance and interpretable spatial outputs essential for real-world moderation workflows.
Continuous advancements of deep learning techniques have profoundly influenced Artificial Intelligence (AI) for visual classification through shifting the field from manual feature engineering to autonomous, hierarchical feature learning. On the contrary, the traditional mechanisms for visual classification relied on various challenges, including large data requirements, computational demands, model interpretability issues, and bias concerns, which severely limited accurate classification. Therefore, the research proposes the Fight-Or-Free Optimized Distributed Patch-Wise Attention-Driven Bidirectional Long Short-Term Memory Network (F2DPAB-Net) for visual classification. The Fight-Or-Free Optimization (F2Opt) algorithm significantly tunes the hyperparameters using stochastic behaviors, potentially improving convergence speed and providing a balance between local exploitation as well as global exploration. Integration of patch-wise triplet attention fusion mechanism offers parallel processing, making the model more efficient in learning long-range dependencies that substantially increase the significance while training. On top of that, utilization of multimodality features enables the model to process and understand different modalities that achieve a more comprehensive interpretation of information and improve generalization. Overall, the proposed F2DPAB-Net outperforms existing methods, thus attaining a maximum of 0.981 Cohen’s Kappa Score, 0.96 MCC, and 0.984 NPV using the COCO dataset, respectively.
Salient object detection (SOD) is a fundamental computer vision task that aims to identify and segment the most visually attractive regions in natural images. While fully supervised methods achieve high accuracy, they rely on costly pixel-wise labels. Scribble annotation offers an efficient alternative with sparse foreground and background strokes, yet it introduces challenges including supervision sparsity, background interference, and boundary ambiguity. This paper studies the SCWSSOD framework for scribble-supervised SOD and proposes a supervised contrastive consistency (SCC) loss to address its limitations. Extensive experiments across six benchmark datasets (PASCAL-S, DUTS-TE, ECSSD, DUT-OMRON, HKU-IS, THUR15K) demonstrate consistent improvements over the baseline. Notably, on the challenging PASCAL-S dataset, our method increases the maximum F-measure from 0.863 to 0.871 and reduces the mean absolute error from 0.122 to 0.113. Both quantitative metrics and qualitative visualizations indicate the effectiveness of our approach in resolving boundary ambiguity and suppressing background noise, suggesting a promising solution for weakly-supervised SOD.
The accurate assessment of movement quality in rehabilitation training is a critical challenge in both clinical and home-based settings. Traditional methods often rely on subjective evaluations or limited sensor data, which can lead to inconsistent and inaccurate assessments. This paper introduces a novel methodology that integrates surface electromyography (sEMG) signals with RGB-D video data to provide a comprehensive, real-time analysis of rehabilitation movement quality. The proposed framework leverages the Manifold-driven Event Forecaster, a sophisticated model that captures the intricate biomechanical and neuromuscular aspects of human motion. The methodology is composed of three primary components: a Nonlinear Constraint Optimizer that aligns multimodal data onto a shared manifold, an Agent-based Temporal Segmenter that partitions data into meaningful temporal segments, and a Probabilistic Quality Predictor that quantifies movement quality with an uncertainty-aware approach. The framework employs advanced strategies such as manifold alignment refinement and uncertainty-aware modeling to enhance robustness and interpretability. These innovations enable the system to make accurate, real-time predictions of movement quality, effectively addressing challenges such as variability in patient performance and sensor noise. Experimental results demonstrate the efficacy of the proposed approach, showing significant improvements in the precision and reliability of movement quality assessments. This comprehensive solution has the potential to transform rehabilitation practices by providing objective, data-driven insights into patient progress and therapy effectiveness, ultimately contributing to improved patient outcomes and more personalized rehabilitation programs.
Cervical cancer is one of the most commonly diagnosed cancers in women worldwide and remains deadly in low and middle income countries. As it grows slowly, it can be prevented if detected early through routine screening. The Pap smear is acknowledged globally as the definitive screening method for cervical cancer. However, classifying cervical cells from Pap smear images is challenging because of the complex morphological changes in cell structures. This paper proposes a deep learning framework with a cosine similarity-based extreme learning machine, enabling faster training and accurate classification of cervical pap cells. It has been observed that scarcity of data and uneven distribution of data among different classes lead to poor performance of a deep learning based multi-class classifier. To overcome these limitations, a hybrid model incorporating an auto-encoder and co-attention mechanism was used to augment data before feeding it in the proposed cosine similarity-based extreme learning machine (CS-ELM) classifier. The hybrid approach is evaluated on the publicly available SIPaKMeD and LBC datasets, and classification accuracies of [Formula: see text]% and [Formula: see text]% are achieved. The results indicate that, the model demonstrates enhanced effectiveness over current leading classifiers in the field of pap cell classification.
This study presents an intelligent framework for English writing generation that integrates contextual semantics with cognitive modeling to improve writing quality and personalization. The proposed system introduces the Contextualized Semantic Composition Network. This network combines syntactic structure modeling with semantic coherence within a hierarchical attention architecture. Through syntax-aware encoding and semantic-context fusion, the model captures long range linguistic dependencies and maintains alignment between grammatical structure and thematic intent. As a result, local expressions remain consistent with global discourse, enabling the generation of coherent and stylistically appropriate text. Based on this framework, a Hierarchical Cognitive Alignment Mechanism is introduced to simulate cognitive processes involved in human writing. This mechanism includes multi-level cognitive tracking, adaptive cognitive modulation, and global consistency control. These components model the interaction among syntax, semantics, and pragmatic intent during text generation. The mechanism ensures smooth transitions between cognitive states and maintains thematic stability, allowing the generated text to achieve both grammatical correctness and natural reasoning. Experimental results on multiple benchmark datasets demonstrate that the proposed framework outperforms existing models in accuracy, fluency, and contextual coherence. The integration of structural modeling and cognitive alignment enhances both adaptability and consistency. These improvements contribute to more effective and personalized English writing support, and provide insights into the integration of linguistic theory, cognitive modeling, and artificial intelligence in educational applications.
The purpose of integrating the visible and infrared images is to generate a new image with a higher perception level for the human visual system. The fusion techniques play a vital role in surveillance, remote sensing, medical imaging, and military use since visible images provide rich texture and color information, whereas infrared images capture thermal data. However, modern fusion methods occasionally fail to achieve the perfect symmetry between noise removal and detail preservation, leading to loss of critical features. The latest infrared and visible image fusion methods are comprehensively discussed in this research and categorized into various categories, such as spatial domain, transform domain, hybrid approaches, and novel deep learning-based methods. Additionally, the impact of objective evaluation measures on various applications and how they are used to analyze fusion performance is discussed. Experimental research demonstrates that deep learning-based fusion techniques have greater flexibility and robustness, while traditional techniques dominate in maintaining the structure. This paper emphasizes the need for adaptive and application-oriented fusion approaches and discusses existing issues and presents future directions toward the improvement of fusion performance.
In this paper, a systematic review of Machine Learning (ML) and Deep Learning (DL) technologies in the field of medicinal plant classification and identification is provided. A review of 104 papers (published from 2018 to 2024) was conducted using the PRISMA framework to evaluate datasets, feature extraction processes, model frameworks, and results. The review reveals significant heterogeneity in data quality, image source, preprocessing techniques, and assessment processes, thereby limiting the comparability of procedures across studies. The most popular models are still based on CNNs. Still, new models are also actively researched: Vision Transformers (ViTs) and Graph Neural Networks (GNNs) have the potential to consider global dependencies and relational structure. These problems have been identified as the unavailability of data to the wider society, non-uniformity in labelling, and the inability to interpret deep models. Overall, CNN-based models, especially hybrid and ensemble models, show the highest consistency in classification accuracy across studies in the reviewed papers and provide evidence of their current prevalence in the classification of medicinal plants. Still, this effect is highly dataset-dependent, underscoring the need for standardized datasets, integrated evaluation procedures, and interpretable AI methods to enable reliable clinical implementation.
One of the most prevalent and deadly cancers is skin cancer, and early diagnosis of the disease has a significant chance of increasing survival. The traditional methods of diagnosis are clinically satisfactory but have limitations such as high costs, prolonged processing time, and inaccessibility in remote areas. The proposed research outlines a novel deep learning model, Extreme Elman Spike Learning-Driven Deep Arctic Belief Neural Network (EESL-DABNN), in particular, to ensure efficient and successful skin lesion segmentation and classification. It is built on multistage processing where sharpening of images by using fast Kuwahara mean filtering is used as the first stage; then histogram equalization and normalization are used to ensure that the intensity of the pixel remains constant. Rotation, flipping, and brightness adjustments augmentation techniques enhance generalization and overcome the imbalance in the dataset. Feature extraction involves the use of a Dual-Stage Attention MobileNetV2 with large-small kernel convolution to learn local and global features. A residual coordinate dual-encoder fusion attention mechanism is also implemented to add more spatial-perception and channel-perception. Segmentation and classification are done using a lightweight EESL-DABNN, optimized through the Arctic Puffin Optimization algorithm for adaptive search space exploration. The experimental outcomes on the ISIC, Melanoma Detection and HAM1000 datasets have accuracies of 99.92%, 99.90%, and 99.94%, respectively, which are more diagnostically reliable at a low cost of computation in clinical applications at scale.
Camouflaged Object Detection (COD) is a difficult task which involves detecting the objects that merge comfortably with the surrounding environment and at first glance no detection can be made of them by the human eye. This challenge is encountered because of contrast and intricate textures. This comparative analysis covers the different models and methods of doing the task of COD and compares the models in different measures including the Mean Absolute Error, [Formula: see text]-measure, structure-measure and enhanced-alignment measure. The method is very handy in the conservation of wildlife, surveillance and medical diagnosis with regards to security. It is a broad overview of the comparison of 35 different models founded on 10 different methodologies. It also opens up new possibilities to venture further in this area due to optimal better extraction and feature selection methods.
This paper proposes an advanced methodology for the automatic detection and spatiotemporal positioning of traffic incidents using video sequence analysis. Traditional traffic monitoring methods often suffer from limitations such as poor accuracy and lack of real time adaptability. To address these challenges, the paper introduces a dynamic and integrated system that combines a dynamic traffic incident detection model (DTIDM) and a spatiotemporal positioning strategy (SPS). The DTIDM processes video sequences to extract relevant features and employs machine learning algorithms for accurate incident detection, while the SPS uses spatiotemporal data to accurately localize the detected incidents. The DTIDM is designed to handle large scale video data and adapt to diverse traffic conditions. It utilizes deep learning techniques for both spatial and temporal feature extraction, ensuring that the system can process real time video inputs efficiently. Furthermore, the SPS incorporates advanced data fusion techniques to integrate spatial and temporal data, which helps in pinpointing the exact location and time of incidents. This strategy ensures precise positioning and enables effective traffic management and response. Experimental results validate the system’s effectiveness, showing significant improvements in detection accuracy and incident localization compared to existing methods. The combination of these components provides a comprehensive solution to enhance traffic safety and management. The paper also identifies future areas for improvement, particularly in optimizing computational efficiency and integrating additional data sources to handle challenging real world scenarios like adverse weather conditions.
The advancement of medical data handling from conventional paper documents to electronic records enabled secure data movement between authenticated legitimate users. While current identity verification algorithms offer unique solutions, they face significant limitations related to data storage scalability, potential privacy breaches, high computational costs, and the lack of standardized protocols. In order to alleviate these constraints, the research proposes a biometric-Blockchain-based authentication Scheme, a Whirlpool Secure Hash-based Biometric Integrated Key Distribution Function (WShBK) for secure data storage and access in a cloud network. The proposed strong cryptographic scheme generates two unique keys derived from the biometric trait of the patient for both encryption and authentication purposes, ensuring strong protection while accessing and storing the data. Furthermore, the advanced encryption standard WShBK (AWShBK) encryption algorithm leverages the strength of a symmetric block cipher and the unique key, offering robust protection against breaches by rendering intercepted data without the correct decryption key. Furthermore, Hybrid biometric-based zero-knowledge proof (HyBZKP) verification offers secure and private transaction validations while sustaining the blockchain integrity. These advancements of the proposed WShBK scheme improve 0.52 encryption rate and 0.53 decryption rate for 250 users analyzed with an attack compared to other cutting-edge models.
Plant diseases have a significant effect on both the amount and quality of agricultural output. The majority of these illnesses have obvious symptoms, and visual examination of the afflicted leaves is the accepted technique for diagnosing plant leaf diseases. New diseases keep emerging as plant structures and agricultural methods change. To prevent the spread of illness and guarantee the healthy growth of crops, these diseases must be identified and classified as soon as possible. Thus, a novel Parallel Convolutional Bidirectional gated recurrent-based framework for Plant Leaf Disease and Severity Classification (PCB-PLDSC) is proposed in this research. In order to balance the dataset by producing more examples for underrepresented classes, image augmentation is done after bilateral filtering for noise reduction. Moreover, sick regions are isolated using a parallel attention-based U-Net (PA-U-Net) model. Adaptive threshold-based local gradient increasing pattern (AT-LGIP), color characteristics, and median binary pattern (MBP) are used to extract features such as texture, color, and spatial patterns. These are categorized using a hybrid parallel convolutional bidirectional gated recurrent (PCB) model, which integrates a bidirectional gated recurrent units (Bi-GRU) model for disease classification with the multi-residual and layer perceptron-based channel attention parallel convolutional neural networks (MRL-CPCNN) model. The impact of each attribute on predictions is shown during the classification phase by SHapley Additive exPlanations (SHAP)-based explainable artificial intelligence (AI). Finally, for detected diseases, the system estimates severity as mild, moderate, or severe to guide effective crop management.
The digital preservation of ancient books requires reliable restoration and enhancement of document images degraded by aging, environmental exposure, and physical damage. We propose an integrated generative framework that reconstructs corrupted regions and improves legibility while preserving structural and semantic fidelity. The core model, Generative Restoration and Enhancement Network (GREN), combines multimodal feature extraction with generative restoration and multiscale enhancement to capture both local details and global layout for high quality reconstruction. To handle spatially non-uniform degradations, we further introduce an Adaptive Restoration Strategy (ARS) that performs spatially aware degradation modeling, semantic guided enhancement, and iterative refinement to produce coherent text and background restoration. Experiments on diverse degradation scenarios demonstrate that the proposed method improves image quality, structural consistency, and readability compared with existing approaches, showing robust performance in restoring historical text and fine textures. The proposed framework provides a practical solution for large scale digital archiving of ancient documents.
Potholes are hazardous defects that occur on the road in the shape of a bowl due to a multitude of factors. Traditionally, enormous image processing and artificial intelligence algorithms are developed to determine the potholes. Yet, their inefficiency in pothole detection by analyzing the images from numerous angles, and their inability to determine the small potholes, reduced their pothole detection accuracy. Hence, an effective multi-head network-based fused attention-enabled Explainable Gated Recurrent Unit-Convolutional Neural Network (MXGRCN) model is proposed to tackle certain hindrances. The usage of the Hawk Eye-based Image enhancement approach analyzes the road images from different angles, which further aids in the better detection of the miniature potholes. Moreover, the integration of the Hybrid Semantic Pixel Transformation method for abstracting the predominant features stipulated the research to determine the potholes better by mitigating the overfitting problems. Furthermore, the amalgamation of the Crocuta Co-operative Hunting Optimization (C2HOpt) and the distributed learning methodology effectively tuned the parameters of the MXGRCN, thereby aiding the optimal convergence attainment with enhanced generalizability. In addition, the utilization of the multi-head attention mechanism enhanced the interpretability of the research with improved model stability. Experimental results show that the proposed MXGRCN model achieves exceptional performance, yielding 99.47% accuracy, 99.48% sensitivity, and 99.46% specificity for 90% of the training on the Pothole Detection dataset.
High Efficiency Video Coding (HEVC) is successful at minimizing the video bitrate, but often creates easily visible artifacts like blocking, ringing, and loss of texture, particularly with low-bitrate conditions. This paper introduces the Hierarchical Feature Fusion and Adaptive Convolutional Embedding Network (H2F-ACEN), which is a two-step architecture aimed at the enhancement of HEVC-decoded frames with limited computational costs. In the first phase, a U-Shaped Encoder-Decoder (USED) block is used to optimize the Quantization Bit Stream (QBS) to silence redundant features and reinforce spatial encoding. The second is the Double-Decker CNN and Hierarchical Embedding Network (HEN), which uses the encoded frames to optimize held decoded frames by means of multi-scale representation learning and residual artifact removal. Quantum-inspired feature extraction layer is a layer that is implemented on classical hardware that provides amplitude-based feature diversification without hardware quantum computation. In order to achieve a better stability in the training process and lower the convergence stagnation, a Triangular Topology Aggregation Optimizer (TTAO) adapts the model parameters of the two stages. The experimental assessments of five benchmark sequences, Lady in Rain, Joy Scene, Rain Fall, Vegetables, and Flower, show that there are substantial objective and perceptual video quality improvements. The average PSNR gain of 2.5 dB and higher SSIM and a MOS score of 4.82 achieved by the proposed model outperform the current techniques like CBT-Net, MBC-LBC, and FVC. These findings make H2F-ACEN a powerful and scalable real-time HEVC video enhancement solution.
