Feature Extraction Network Research Articles

I-DINO: High-Quality Object Detection for Indoor Scenes

Object Detection in Complex Indoor Scenes is designed to identify and categorize objects in indoor settings, with applications in areas such as smart homes, security surveillance, and home service robots. It forms the basis for advanced visual tasks including visual question answering, video description generation, and instance segmentation. Nonetheless, the task faces substantial hurdles due to background clutter, overlapping objects, and significant size differences. To tackle these challenges, this study introduces an indoor object detection approach utilizing an enhanced DINO framework. To cater to the needs of indoor object detection, an Indoor-COCO dataset was developed from the COCO object detection dataset. The model incorporates an advanced Res2net as the backbone feature extraction network, complemented by a deformable attention mechanism to better capture detailed object features. An upgraded Bi-FPN module is employed to replace the conventional feature fusion module, and SIoU loss is utilized to expedite convergence. The experimental outcomes indicate that the refined model attains an mAP of 62.3%, marking a 5.2% improvement over the baseline model. These findings illustrate that the DINO-based indoor object detection model exhibits robust generalization abilities and practical utility for multi-scale object detection in complex environments.

Classification of hand movements from EEG using a FusionNet based LSTM network

Abstract Objective.
Accurate classification of electroencephalogram (EEG) signals is crucial for advancing brain-computer interface (BCI) technology. However, current methods face significant challenges in classifying hand movement EEG signals, including effective spatial feature extraction, capturing temporal dependencies, and representing underlying signal dynamics.
Approach.
This paper introduces a novel multi-model fusion approach, FusionNet-LSTM, designed to address these issues. Specifically, it integrates Convolutional Neural Networks (CNN) for spatial feature extraction, Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies, and Autoregressive (AR) models for representing signal dynamics.
Main results.
Compared to single models and state-of-the-art methods, this fusion approach demonstrates substantial improvements in classification accuracy. Experimental results show that the proposed model achieves an accuracy of 87.1% in cross-subject data classification and 99.1% in within-subject data classification. Additionally, Gradient Boosting Trees were employed to evaluate the significance of various EEG features to the model.
Significance.
This study highlights the advantages of integrating multiple models and introduces a superior classification model, which is pivotal for the advancement of BCI systems.

Outpatient scheduling problem in smart hospital with two-agent deep reinforcement learning algorithm

As the internet and digital technologies revolutionize the healthcare sector, the development of smart hospitals is pivotal for improving medical service efficiency. This research presents a mathematical model for the outpatient scheduling problem and introduces the Graph-based Two-Agent Deep Reinforcement Learning algorithm (GTARS-DRL) to address the escalating demand for medical services and the scarcity of resources. The GTARS-DRL leverages graph neural networks for feature extraction of patient medical activities and applies the PDQPN algorithm to train agents for optimizing patient sequencing and service desk allocation, effectively minimizing wait times. Our experiments indicate that GTARS-DRL outperforms traditional scheduling rules in solution quality and matches the performance of Genetic Algorithms (GA), with a notable advantage in computational efficiency, boasting an ACPU of 0.762 s and an ARPT of 0.327 s. The algorithm’s efficacy is further supported by its successful application to larger-scale instances, showcasing its potential for real-world use. This study contributes a novel approach to outpatient scheduling in smart hospitals, providing theoretical backing and practical direction for the optimization of medical resource allocation, and is poised to significantly enhance service efficiency and patient satisfaction.

Text Detection on Industrial Barrel Label with Convolutional Attention and Dual‐Branch Feature Network

Industrial barrel labels generally have low visual contrast, uneven lighting, and cluttered background, making it challenging to accurately locate text regions. This paper proposes a text detection network to solve the inaccurate localization problem based on DBNet. First, a convolutional attention mechanism is applied to the feature extraction network to get more valuable text feature maps. Then, a dual‐branch convolutional feature module is proposed in the feature pyramid to enrich contextual information. Besides, during the probability map generation stage, using a feature remodeling enhancement module to further distinguish text and text boundaries. This paper designs comparative experiments on ILTD, ICDAR2015 and MSRA‐TD500 datasets, achieve F‐measure of 92.3%, 86.0% and 84.1%, which are 2.2%, 2.3%, and 1.9% higher than DBNet, respectively. They demonstrate that our proposed method exhibits competitive performance and strong robustness. © 2024 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.

CSTAN: A Deepfake Detection Network with CST Attention for Superior Generalization

With the advancement of deepfake forgery technology, highly realistic fake faces have posed serious security risks to sensor-based facial recognition systems. Recent deepfake detection models mainly use binary classification models based on deep learning. Despite achieving high detection accuracy on intra-datasets, these models lack generalization ability when applied to cross-datasets. We propose a deepfake detection model named Channel-Spatial-Triplet Attention Network (CSTAN), which focuses on the difference between real and fake features, thereby enhancing the generality of the detection model. To enhance the feature-learning ability of the model for image forgery regions, we have designed the Channel-Spatial-Triplet (CST) attention mechanism, which extracts subtle local information by capturing feature channels and the spatial correlation of three different scales. Additionally, we propose a novel feature extraction method, OD-ResNet-34, by embedding ODConv into the feature extraction network to enhance its dynamic adaptability to data features. Trained on the FF++ dataset and tested on the Celeb-DF-v1 and Celeb-DF-v2 datasets, the experimental results show that our model has stronger generalization ability in cross-datasets than similar models.

A novel digital-twin approach based on transformer for photovoltaic power prediction

The prediction of photovoltaic (PV) system performance has been intensively studied as it plays an important role in the context of sustainability and renewable energy generation. In this paper, a digital twin (DT) model based on a domain-matched transformer is proposed using convolutional neural network (CNN) for domain-invariant feature extraction, transformer for PV performance prediction, and domain adaptation neural network (DANN) for domain adaptation. The effectiveness of the proposed framework is validated using a PV power prediction dataset. The results indicate an accuracy improvement of up to 39.99% in model performance. Additionally, experiments with varying numbers of timestamps demonstrate enhanced PV power prediction performance as parameters are continuously updated within the DT framework, offering a reliable solution for real-time and adaptive PV power forecasting.

A novel adjunctive diagnostic method for bone cancer: Osteosarcoma cell segmentation based on Twin Swin Transformer with multi-scale feature fusion

BackgroundOsteosarcoma, the most common primary bone tumor originating from osteoblasts, poses a significant challenge in medical practice, particularly among adolescents. Conventional diagnostic methods heavily rely on manual analysis of magnetic resonance imaging (MRI) scans, which often fall short in providing accurate and timely diagnosis. This underscores the critical need for advancements in medical imaging technologies to improve the detection and characterization of osteosarcoma. MethodsIn this study, we sought to address the limitations of current diagnostic approaches by leveraging Hoechst-stained images of osteosarcoma cells obtained via fluorescence microscopy. Our primary objective was to enhance the segmentation of osteosarcoma cells, a crucial step in precise diagnosis and treatment planning. Recognizing the shortcomings of existing feature extraction networks in capturing detailed cellular structures, we propose a novel approach utilizing a twin swin transformer architecture for osteosarcoma cell segmentation, with a focus on multi-scale feature fusion. ResultsThe experimental findings demonstrate the effectiveness of the proposed Twin Swin Transformer with multi-scale feature fusion in significantly improving osteosarcoma cell segmentation. Compared to conventional techniques, our method achieves superior segmentation performance, highlighting its potential utility in clinical settings. ConclusionThe development of our Twin Swin Transformer with multi-scale feature fusion method represents a significant advancement in medical imaging technology, particularly in the field of osteosarcoma diagnosis. By harnessing advanced computational techniques and leveraging high-resolution imaging data, our approach offers enhanced accuracy and efficiency in osteosarcoma cell segmentation, ultimately facilitating better patient care and clinical decision-making.

Green apple detection method based on multi-dimensional feature extraction network model and Transformer module

To enhance fast and accurate detection of pollution-free green apples for food safety, this paper uses the DETR network as a framework to propose a new method for pollution-free green apple detection based on a multi-dimensional feature extraction network and Transformer module. Firstly, an improved DETR network main feature extraction module adopts the ResNet18 network and replaces some residual layers with deformable convolutions (DCNv2), enabling the model to better adapt to pollution-free fruit changes at different scales and angles, while eliminating the impact of microbial contamination on fruit testing; Subsequently, the extended spatial pyramid pooling model (DSPP) and multi-scale residual aggregation module (FRAM) are integrated, which help reduce feature noise and minimize the loss of underlying features during the feature extraction process. The fusion of the two modules enhances the model’s ability to detect objects of different scales, thereby improving the accuracy of near-color fruit detection; At the same time, in order to solve the problems of slow convergence speed and large calculation amount of the basic network model, the convergence speed of the overall network model is improved by replacing the attention mechanism of Transformer. Experimental results show that compared with the original DETR model, the proposed algorithm has improved in AP, AP50 and AP75 indicators, especially in the AP50 indicator, which has the most obvious improvement reaching a detection accuracy of 97.12%. In the meantime, the trained network model is deployed on the picking robot. Compared with the original DETR network model, its average detection accuracy is as high as 96.58%, and the detection speed is increased by about 51%. Mixed sample detection tests were carried out before and after the model deployment, and the detection rate of the proposed method for non-polluted fruits reached more than 0.95. enabling the picking robot to efficiently complete the task of picking green apples. The test results show that the algorithm proposed in this article exhibits great potential in the task of detecting pollution-free near-color fruits by the picking robot. It ensures pollution-free fruit picking and the application of AI in food safety.

A Variable-Iterative Fully Convolutional Neural Network for Sparse Unmixing

Neural networks have greatly promoted the development of hyperspectral unmixing (HU). Most data-driven deep networks extract features of hyperspectral images (HSIs) by stacking convolutional layers to achieve endmember extraction and abundance estimation. Some model-driven networks have strong interpretability but fail to mine the deep feature. We propose a variable-iterative fully convolutional neural network (VIFCNN) for sparse unmixing, combining the characteristics of these two networks. Under the model-driven iterative framework guided by sparse unmixing by variable splitting and augmented lagrangian (SUnSAL), a data-driven spatial-spectral feature learning module and a spatial information updating module are introduced to enhance the learning of data information. Experimental results on synthetic and real datasets show that VIFCNN significantly outperforms several traditional unmixing methods and two deep learning???based methods. On real datasets, our method improves signal-to-reconstruction error by 17.38%, reduces abundance root-mean-square error by 25.24%, and reduces abundance spectral angle distance by 31.40% compared with U-ADMM-ßUNet.

A multimodal educational robots driven via dynamic attention

IntroductionWith the development of artificial intelligence and robotics technology, the application of educational robots in teaching is becoming increasingly popular. However, effectively evaluating and optimizing multimodal educational robots remains a challenge.MethodsThis study introduces Res-ALBEF, a multimodal educational robot framework driven by dynamic attention. Res-ALBEF enhances the ALBEF (Align Before Fuse) method by incorporating residual connections to align visual and textual data more effectively before fusion. In addition, the model integrates a VGG19-based convolutional network for image feature extraction and utilizes a dynamic attention mechanism to dynamically focus on relevant parts of multimodal inputs. Our model was trained using a diverse dataset consisting of 50,000 multimodal educational instances, covering a variety of subjects and instructional content.Results and discussionThe evaluation on an independent validation set of 10,000 samples demonstrated significant performance improvements: the model achieved an overall accuracy of 97.38% in educational content recognition. These results highlight the model's ability to improve alignment and fusion of multimodal information, making it a robust solution for multimodal educational robots.

Lightweight and efficient deep learning models for fruit detection in orchards

The accurate recognition of apples in complex orchard environments is a fundamental aspect of the operation of automated picking equipment. This paper aims to investigate the influence of dense targets, occlusion, and the natural environment in practical application scenarios. To this end, it constructs a fruit dataset containing different scenarios and proposes a real-time lightweight detection network, ELD(Efficient Lightweight object Detector). The EGSS(Efficient Ghost-shuffle Slim module) module and MCAttention(Mix channel Attention) are proposed as innovative solutions to the problems of feature extraction and classification. The attention mechanism is employed to construct a novel feature extraction network, which effectively utilizes the low-latitude feature information, significantly enhances the fine-grained feature information and gradient flow of the model, and improves the model’s anti-interference ability. Eliminate redundant channels with SlimPAN to further compress the network and optimise functionality. The network as a whole employs the Shape-IOU loss function, which considers the influence of the bounding box itself, thereby enhancing the robustness of the model. Finally, the target detection accuracy is enhanced through the transfer of knowledge from the teacher’s network through knowledge distillation, while ensuring that the overall network is sufficiently lightweight. The experimental results demonstrate that the ELD network, designed for fruit detection, achieves an accuracy of 87.4%. It has a relatively low number of parameters (4.3×105\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$4.3 \ imes 10^5$$\\end{document}), a GLOPs of only 1.7, and a high FPS of 156. This network can achieve high accuracy while consuming fewer computational resources and performing better than other networks.

Multi‐scale feature extraction for energy‐efficient object detection in remote sensing images

AbstractObject detection in remote sensing images aims to interpret images to obtain information on the category and location of potential targets, which is of great importance in traffic detection, marine supervision, and space reconnaissance. However, the complex backgrounds and large scale variations in remote sensing images present significant challenges. Traditional methods relied mainly on image filtering or feature descriptor methods to extract features, resulting in underperformance. Deep learning methods, especially one‐stage detectors, for example, the Real‐Time Object Detector (RTMDet) offers advanced solutions with efficient network architectures. Nevertheless, difficulty in feature extraction from complex backgrounds and target localisation in scale variations images limits detection accuracy. In this paper, an improved detector based on RTMDet, called the Multi‐Scale Feature Extraction‐assist RTMDet (MRTMDet), is proposed which address limitations through enhancement feature extraction and fusion networks. At the core of MRTMDet is a new backbone network MobileViT++ and a feature fusion network SFC‐FPN, which enhances the model's ability to capture global and multi‐scale features by carefully designing a hybrid feature processing unit of CNN and a transformer based on vision transformer (ViT) and poly‐scale convolution (PSConv), respectively. The experiment in DIOR‐R demonstrated that MRTMDet achieves competitive performance of 62.2% mAP, balancing precision with a lightweight design.

A method for recognizing abnormal behaviors of personnel at petroleum stations based on GTB-ResNet

PurposeThe abnormal behaviors of staff at petroleum stations pose significant safety hazards. Addressing the challenges of high parameter counts, lengthy training periods and low recognition rates in existing 3D ResNet behavior recognition models, this paper proposes GTB-ResNet, a network designed to detect abnormal behaviors in petroleum station staff.Design/methodology/approachFirstly, to mitigate the issues of excessive parameters and computational complexity in 3D ResNet, a lightweight residual convolution module called the Ghost residual module (GhostNet) is introduced in the feature extraction network. Ghost convolution replaces standard convolution, reducing model parameters while preserving multi-scale feature extraction capabilities. Secondly, to enhance the model's focus on salient features amidst wide surveillance ranges and small target objects, the triplet attention mechanism module is integrated to facilitate spatial and channel information interaction. Lastly, to address the challenge of short time-series features leading to misjudgments in similar actions, a bidirectional gated recurrent network is added to the feature extraction backbone network. This ensures the extraction of key long time-series features, thereby improving feature extraction accuracy.FindingsThe experimental setup encompasses four behavior types: illegal phone answering, smoking, falling (abnormal) and touching the face (normal), comprising a total of 892 videos. Experimental results showcase GTB-ResNet achieving a recognition accuracy of 96.7% with a model parameter count of 4.46 M and a computational complexity of 3.898 G. This represents a 4.4% improvement over 3D ResNet, with reductions of 90.4% in parameters and 61.5% in computational complexity.Originality/valueSpecifically designed for edge devices in oil stations, the 3D ResNet network is tailored for real-time action prediction. To address the challenges posed by the large number of parameters in 3D ResNet networks and the difficulties in deployment on edge devices, a lightweight residual module based on ghost convolution is developed. Additionally, to tackle the issue of low detection accuracy of behaviors amidst the noisy environment of petroleum stations, a triple attention mechanism is introduced during feature extraction to enhance focus on salient features. Moreover, to overcome the potential for misjudgments arising from the similarity of actions, a Bi-GRU model is introduced to enhance the extraction of key long-term features.

AGCNAF: predicting disease-gene associations using GCN and multi-head attention to fuse the similarity features

Abstract Association prediction between diseases and genes is a critical step in revealing the molecular mechanisms of diseases and developing drug treatment strategies. With the explosive growth of data in the biomedical field, how to effectively utilize these data for accurate prediction has become a hotspot and challenge in current research. To overcome the limitations of current prediction methods in dealing with complex biological network structures and feature extraction, this study proposes AGCNAF, a method that combines an unsupervised Graph Convolutional Network (GCN) and a multi-head attention mechanism. The metagraph-guided random walk strategy enables AGCNAF to capture local and high-order topological structures in the graph, while GCN is responsible for realizing deep feature extraction of these structures. By incorporating similarity features through the multi-attention mechanism, AGCNAF achieves effective integration of global and local features, which significantly improves the prediction performance. By incorporating similarity features through the multi-attention mechanism, AGCNAF achieves effective integration of global and local features, which significantly improves the prediction performance. By utilizing the machine learning binary classification model for prediction, the experimental results through five-fold cross-validation show that AGCNAF demonstrates significant advantages in prediction performance compared to existing methods, with its AUC and AUPR reaching 0.9686 and 0.9709, respectively, and the AUC up to 0.9812 under specific conditions. 
To verify the practical application value of AGCNAF, this study also conduct case studies on Alzheimer's disease, lung cancer, and breast cancer. The results further confirm the excellent performance of AGCNAF in identifying potential disease-gene associations, which opens up new possibilities for future disease-gene research.

Research on Deep Learning Detection Model for Pedestrian Objects in Complex Scenes Based on Improved YOLOv7

Objective: Pedestrian detection is very important for the environment perception and safety action of intelligent robots and autonomous driving, and is the key to ensuring the safe action of intelligent robots and auto assisted driving. Methods: In response to the characteristics of pedestrian objects occupying a small image area, diverse poses, complex scenes and severe occlusion, this paper proposes an improved pedestrian object detection method based on the YOLOv7 model, which adopts the Convolutional Block Attention Module (CBAM) attention mechanism and Deformable ConvNets v2 (DCNv2) in the two Efficient Layer Aggregation Network (ELAN) modules of the backbone feature extraction network. In addition, the detection head is replaced with a Dynamic Head (DyHead) detector head with an attention mechanism; unnecessary background information around the pedestrian object is also effectively excluded, making the model learn more concentrated feature representations. Results: Compared with the original model, the log-average miss rate of the improved YOLOv7 model is significantly reduced in both the Citypersons dataset and the INRIA dataset. Conclusions: The improved YOLOv7 model proposed in this paper achieved good performance improvement in different pedestrian detection problems. The research in this paper has important reference significance for pedestrian detection in complex scenes such as small, occluded and overlapping objects.

Advancing Ton-Bag Detection in Seaport Logistics with an Enhanced YOLOv8 Algorithm

Intelligent logistics and freight transportation is an important part of realizing the intelligence of port terminals. Due to the problems of inaccurate ton bag identification, high costs, large model sizes, and long computation times in traditional freight transportation—issues that hinder meeting real-time requirements on resource-constrained operational equipment—this paper proposes an improved lightweight ton bag detection algorithm, YOLOv8-TB (YOLOv8-Ton Bag), which is optimized based on YOLOv8. Firstly, the improved LZKAC module is introduced to combine with SPPF to form a new SPPFLKZ module, which improves the feature expression performance. Then, with reference to spatial and channel reconstruction convolution and deformable convolution, the C2f-SCTT block is designed for the backbone network, which reduces the spatial and channel redundancy between features in the network. Finally, the C2f-ORECZ block based on a linear scaling layer is designed for the neck, which reduces the training overhead and strengthens the feature learning of the feature extraction network for the targets in the complex background of the harbor and adds the 160 × 160 scale detection head to strengthen small target detection abilities. On the logistics ton bag operation dataset provided by shipping port enterprises, the improved algorithm improves by 3.7% and 5% compared with the original algorithm in mAP50 and mAP50-95, respectively, the model size is reduced by 4.42 MB and the amount of model computation is only 8 G, which is capable of accurately detecting logistics ton bags in real time. The superiority of the method is verified by comparing it with other classical target detection algorithms.

Text-Enhanced Graph Attention Hashing for Cross-Modal Retrieval

Deep hashing technology, known for its low-cost storage and rapid retrieval, has become a focal point in cross-modal retrieval research as multimodal data continue to grow. However, existing supervised methods often overlook noisy labels and multiscale features in different modal datasets, leading to higher information entropy in the generated hash codes and features, which reduces retrieval performance. The variation in text annotation information across datasets further increases the information entropy during text feature extraction, resulting in suboptimal outcomes. Consequently, reducing the information entropy in text feature extraction, supplementing text feature information, and enhancing the retrieval efficiency of large-scale media data are critical challenges in cross-modal retrieval research. To tackle these, this paper introduces the Text-Enhanced Graph Attention Hashing for Cross-Modal Retrieval (TEGAH) framework. TEGAH incorporates a deep text feature extraction network and a multiscale label region fusion network to minimize information entropy and optimize feature extraction. Additionally, a Graph-Attention-based modal feature fusion network is designed to efficiently integrate multimodal information, enhance the affinity of the network for different modes, and retain more semantic information. Extensive experiments on three multilabel datasets demonstrate that the TEGAH framework significantly outperforms state-of-the-art cross-modal hashing methods.

Heart sound classification algorithm based on bispectral feature extraction and convolutional neural networks

Cardiovascular disease (CVD) is one of the leading causes of death worldwide. Heart sound classification plays a key role in the early detection of CVD. The difference between normal and abnormal heart sounds is not obvious. In this paper, in order to improve the accuracy of the heart sound classification model, we propose a heart sound feature extraction method based on bispectral analysis and combine it with convolutional neural network (CNN) to classify heart sounds. The model can effectively suppress Gaussian noise by using bispectral analysis and can effectively extract the features of heart sound signals without relying on the accurate segmentation of heart sound signals. At the same time, the model combines with the strong classification performance of convolutional neural network and finally achieves the accurate classification of heart sound. According to the experimental results, the proposed algorithm achieves 0.910, 0.884 and 0.940 in terms of accuracy, sensitivity and specificity under the same data and experimental conditions, respectively. Compared with other heart sound classification algorithms, the proposed algorithm shows a significant improvement and strong robustness and generalization ability, so it is expected to be applied to the auxiliary detection of congenital heart disease.

Identifying rice lodging based on semantic segmentation architecture optimization with UAV remote sensing imaging

Lodging, a prevalent issue during rice growth, detrimentally impacts both yield and quality. It also complicates the harvesting process, reducing the efficiency of mechanized collection. Existing monitoring methods, predominantly based on manual observation and satellite remote sensing, fall short in addressing the requirements of contemporary, efficient, and real-time agriculture. This research integrates image analysis techniques with advanced optimization algorithms to develop a semantic segmentation model specifically designed for detecting rice lodging in remote sensing images. The model, named MI-UConvNeXt, employs a ConvNeXt-based feature extraction network utilizing UNet architecture (UConvNeXt) and incorporates an improved multi-objective salp swarm algorithm with Latin hypercube sampling and an elite opposition-based learning strategy (ISSA-LE) to dynamically adjusting the number of UConvNeXt channels. MI-UConvNeXt achieves a balance between accuracy and complexity. Compared to seven other semantic segmentation models from the literature, MI-UConvNeXt exhibits enhanced performance, with a Pixel Accuracy (PA) of 95.59%, mean Pixel Accuracy (mPA) of 95.62%, and mean Intersection over Union (mIoU) of 91.91% on the validation set. This demonstrates the model’s superior accuracy, lower computational resource demands, and enhanced efficiency. By integrating deep learning with intelligent optimization algorithms, this study offers a novel and effective approach for monitoring crop lodging in agricultural production, providing robust technical support for the accurate extraction of crop phenotypic information.

An attention fused sequence -to-sequence convolutional neural network for accurate solar irradiance forecasting and prediction using sky images

The challenge of accurately forecasting ultra-short-term solar irradiance for photovoltaic systems is complicated by rapidly changing weather, and while ground-based sky images offer potential improvements, effectively extracting spatiotemporal data from these images remains a significant hurdle for current computer vision models. A new hybrid model, "An Attention Fused Sequence-to-Sequence Convolutional Neural Network," is being developed to address this challenge. The model predicts intra-hour GHI, DNI, and DHI with a 10-min lead time by combining a Convolutional Neural Network (for spatial feature extraction from sky images), an attention mechanism (to focus on relevant regions), and a sequence-to-sequence model (for temporal feature extraction from time-series data). The proposed Model is trained using the NREL Solar Radiation Research Laboratory Dataset while evaluating the model with the Mean Bias Error, Mean Absolute Error, Root Mean Squared Error, R Squared, and forecasting skill score. The 10-min and sequence length 2 interval is considered to be the best performing across most of the evaluation metrics with an MBE value is 2.321 W/m2, MAE value of 39.490 W/m2, RMSE value of 62.086 W/m2, R2 value is 0.909 W/m2, FSS value of 23.589 W/m2 and an MBE of 4.876 W/m2, MAE of 56.887 W/m2, RMSE of 85.346 W/m2, R2 of 0.834 and FSS of 24.881 W/m2 respectively. Furthermore, the sensitivity analysis reveals that the proposed model's performance is influenced by both the sequence length and the lead time. The proposed framework outperforms other techniques for ultra-short-term PV generation forecasting, demonstrating its potential for practical deployment in PV systems to improve grid reliability and energy management.