Driven by advancements in information technology, remote education has rapidly emerged as a flexible learning model that transcends time and space constraints.Behavior monitoring plays a vital role in ensuring the quality of remote instruction.However, the complexity of remote learning environments-marked by variations in lighting conditions and scene differences-poses significant challenges for accurate behavior monitoring based on multimodal image data.Existing multimodal image fusion methods often fail to effectively utilize deep-level features, while deep learning-based approaches exhibit limited capacity for adaptive fusion in complex scenarios.Furthermore, conventional data augmentation techniques generally lack task-specific strategies tailored for behavior monitoring in remote education, and methods such as generative adversarial networks (GANs) suffer from issues like mode collapse and suboptimal performance in multimodal data augmentation.This paper addresses the challenge of adaptive enhancement in multimodal image fusion for behavior monitoring in remote education.We propose a diffusion model-based multimodal image generation algorithm that extracts latent features across different modalities to synthesize high-quality fused data, mitigating data scarcity and quality issues.Additionally, we introduce a task-oriented adaptive enhancement method that dynamically optimizes augmentation strategies based on the learning context and monitoring requirements, thereby improving data diversity and model adaptability.The proposed framework provides more accurate data support for remote education behavior monitoring, significantly enhancing the generalization and robustness of monitoring models.These findings offer theoretical and practical value for personalized education and the advancement of multimodal data processing technologies.

Emotion recognition from audio data holds immense potential in revolutionizing humancomputer interaction (HMI), affective computing, and psychological health monitoring.This paper delves into a novel deep learning approach that leverages the strengths of multimodal features mined from audio signals.We propose a model that transcends the disadvantages of existing methods by combining Mel-Frequency Cepstral Coefficients (MFCCs) with high-level representations extracted from a pre-trained DenseNet architecture.MFCCs provide a compressed representation of the audio signal's spectral characteristics, capturing crucial emotional cues like pitch and intensity.These learned patterns can translate to the domain of audio emotion recognition, enabling the model to identify subtle emotional nuances that might be difficult to capture with traditional feature engineering techniques.Our deep learning model, comprised of dense layers, fosters robust performance in accurately classifying emotions across diverse categories.We used a Melspectrograms-based LSTM model for speech emotion recognition that effectively identifies various emotions.We rigorously evaluate the proposed approach on the TESS dataset.The experimental results are truly compelling, showcasing a staggering accuracy of 100%.This exceptional performance signifies the effectiveness of the multimodal approach in extracting and interpreting emotional cues from audio data.

Hybrid classifiers are neural network models that are structured with two different network architectures to create the best deep learning model.The hybrid models are developed generally to meet certain complex feature learning problems.However, the nature of hybrid classifiers causes the network nodes to learn 'n' number of features, which increases the computational complexity of the classifiers.Therefore, the implementation of a hybrid classifier reaches an impossible state when the processor and memory spaces are restricted in real-time applications.The proposed work utilizes the feedforward neural network architecture with two hidden layers in combination with the eXtreme Gradient Boosting (XGBoost) algorithm to form a hybrid classifier called H-StrokeNet, and a Crow Search Optimization (CSO) algorithm is included in the work to reduce the complexity of the hybrid network by providing optimized feature extraction.The performance of the proposed work is verified with the regular XGBoost algorithm and deep neural networks using the Laboratory of Processing Image, Signals, and Computer Science (LAPISCO) dataset, and a comparative analysis is made with different algorithms such as support vector machine (SVM), Decision Tree (DT), Na ve Bayes (NB), k-Nearest Neighbor (kNN), and XGBoost.Results show the superior performance of the proposed system with an overall accuracy of 96.55%, whereas it is 92.09% with the XGBoost classifier.

Cancer ranks among the top causes of death, primarily due to its tendency to be diagnosed at a late stage and its highly aggressive nature.Lung cancer, in particular, has one of the highest mortality rates due to delayed diagnosis.The purpose of this study is to address this issue by applying cutting-edge deep learning methods for the prompt identification of lung cancer from CT scan images.In particular, various Convolutional Neural Network (CNN) architectures are utilized to distinguish between cancerous and non-cancerous CT scan images.To further enhance classification accuracy, a spatial attention mechanism is integrated into the models, enabling them to focus on relevant image regions.The developed models are assessed using a publicly accessible dataset and assessed based on various performance indicators, including precision, recall, accuracy and the F1-score.The findings highlight the superior performance of the attention-enhanced DenseNet model, achieving significant improvements in classification metrics compared to other architectures.The research supports the improvement of reliable computer-aided diagnosis systems for lung cancer, with a focus on the effectiveness of embedding attention modules within CNNs for medical imaging tasks.

Although Turkey ranks first in the world in terms of raisin exports, it ranks fourth in revenue.The main reason for this problem is that the quality of exported raisins does not meet the desired level.In this study, an intelligent real-time system was developed to determine suitability for export by classifying raisins according to color and capstem amount according to the criteria of TS 3411.A dataset of 2336 Sultana raisins was created and expanded to 10,544 images using data augmentation techniques such as exposure adjustment, rotation, and noise addition.These techniques not only increased the dataset size but also significantly improved the model performance, as evidenced by the remarkable success of the CNN models with an average mAP of 98.8%.The images were evaluated in real time in the developed GUI software and their compliance with the standard was determined.The study is the first of its kind to classify and detect raisins according to the standard TS 3411 using an intelligent real-time system.

Pap smears, also known as pap tests, can identify abnormal cells early that develop cervical cancer, allowing for timely intervention and treatment.Even though the incidence rate is reduced in this modern era, it poses a significant risk to human life and should be taken very seriously.An accurate and rapid system for classifying pap smear images is necessary to provide appropriate therapy.Deep Neural Networks (DNNs) have garnered much interest in recent years and have shown outstanding categorization results in computer vision.An efficient Pap Smear Image Classification (PSIC) system with Efficient Weight Regularization (EWR) in DNN is presented in this study.The main problem with neural networks is that they have large weights that overfit the training data.To overcome this difficulty, the EWR approach is employed to penalize the large weights using grid search.The proposed non-invasive support system detects pap-smear images with cancerous cells.The HERLEV dataset comprises 675 digitized abnormal images, and 242 normal images are utilized for the classification task.When using the capabilities of the EWR-DNN combination, the proposed PSIC system can work at its absolute best.The concepts described in this study also provide a possible path to increase the categorization accuracy of all medical diagnoses.Results show that the PSIC system, which employs EWR approach achieves 98.9% classification accuracy, 99.3% specificity and 98.5% sensitivity using a regularization parameter of 10 -3 .The comparison study with other deep learning models such as VGG, ResNet, AlexNet and GoogleNet also shows the superior performance of the PSIC system.

This work focuses on using level set curves for medical image segmentation through the DRLSE (Distance Regularization Level Set Evolution) algorithm, recognized for its effectiveness and adaptability.Traditional systems face limitations in computation time and efficiency when implementing this algorithm.To overcome these challenges, FPGA (Field-Programmable Gate Arrays) are used for their parallelism and low resource consumption.The objective is to optimize medical image segmentation by implementing the DRLSE algorithm on FPGA while ensuring efficient resource and computation time management.The Algorithm was first simulated in MATLAB and tested on a database of brain, breast, and other medical images, demonstrating its robustness and flexibility.The results validate the effectiveness of the DRLSE algorithm and highlight the advantages of the FPGA in terms of speed and precision.Despite the limited documentation on implementing DRLSE on FPGA Our approach is distinguished by the use of DDR memory, which provides increased capacity to overcome the limitations of BRAM memory.Parameter optimization ensures better performance and efficient management of hardware resources.This work underscores the potential of FPGA-based implementations for accelerating computationally intensive tasks like medical image segmentation while maintaining high accuracy and efficiency.

With the significant growth of malicious attacks, safeguarding personal data has become a critical and pressing concern.Images transmitted over unsecured networks are particularly vulnerable to tampering and unauthorized distribution.To address these challenges, networks must be fortified with robust strategies capable of preventing new attacks.These strategies should prioritize enhanced performance while maintaining content fidelity.The article presents a blind approach for color image watermarking that leverages the YCbCr color space's key properties.Both the image and the watermark were converted from RGB to YCbCr.Afterward, the watermark is encrypted using the Arnold Chaotic Map (ACM) to strengthen privacy and embedded into the Cb component using the WH-SVD-Cb watermarking schema.Image authenticity is validated through blind watermark extraction.The method's robustness and imperceptibility are analyzed through empirical evaluations.The findings indicate that embedding of watermark in the Cb component achieves notable robustness with NC values closing 1 and imperceptibility exceeding 52 dB, all within a processing time of 0.224 seconds.