Multimedia Tools and Applications最新文献

In this paper, we propose an optical image encryption scheme based on modified 3D double-phase encoding algorithm (3D-DPEA) in the gyrator transform (GT) domain, in which a plaintext is encrypted into two sparse volumetric ciphertexts under the constraints of chaos-generated binary amplitude masks (BAMs). Then, the two volumetric ciphertexts are multiplexed into the corresponding 2D ciphertexts for convenient storage and transmission. First, due to the synergistic adjustment of the two sparse volumetric ciphertexts during the iterative process, the 3D-DPEA would achieve higher recovery quality of the decrypted image with fewer iterations. In addition, because the BAMs are generated by the logistic-tent (LT) chaotic map which is closely related to the rotation angles of GT, and the LT chaotic map has several advantages such as nonlinear, pseudorandom behavior, and high sensitivity of initial conditions, the sensitivity of the secret key could be significantly improved by several orders of magnitude, reaching up to 10⁻¹⁴. As a result, the 3D-DPEA scheme not only eliminates the explicit/linear relationship between the plaintext and the ciphertext but also substantially enhances security. For decryption, the corresponding decrypted image can be achieved by recording an intensity pattern when a coherent beam crosses two sparse volumetric ciphertexts sequentially. Furthermore, BAMs wouldn’t impose an additional burden on the storage and transmission of secret keys. A series of numerical simulations are performed to verify the effectiveness and security of the proposed encryption scheme.

The main purpose of indoor localization is to precisely locate users and help them navigate within an indoor area, like a building or campus, where GPS and other satellite technologies lack precision. Our methodology for achieving indoor localization has been to implement classifiers that use Received Signal Strength Indicator (RSSI) values of WiFi signals collected from smart hand-held devices. However, these RSSI values keep varying, often appreciably, from time to time and device to device. So, to instill more generalizability into the location prediction process, ensemble models have been built that can learn from the pros and cons of all of their member classifiers. In this paper, we have presented several neural network based ensemble models to compensate for the lack of detailed studies with ensemble models (especially neural network based ones) on indoor localization. Our second contribution lies in designing a knowledge distillation framework for the ensemble models that preserves the classification performance while make the system real-time responsive as the lightweight distilled model could be executed locally on the edge devices. Our proposed knowledge distillation framework distils the knowledge of a large neural network based ensemble classifier into a much smaller compressed classification model while maintaining the performance. We have implemented and shown the workings of the proposed knowledge distillation framework on three publicly available benchmark datasets. The proposed model have been found to achieve 83.95%, 93.10% and 96.48% accuracy for DataSet1, DataSet2 and DataSet3, respectively.

Accurately predicting the risk of diabetes is of paramount importance for early intervention and prevention. To achieve precise diabetes risk prediction, we propose a hybrid diabetes risk prediction model, XGB-ILSO-1DCNN, which combines the Extreme Gradient Boosting (XGBoost) algorithm, the Improved Lion Swarm Optimization algorithm, and the deep learning model 1DCNN. Firstly, an XGBoost is trained based on the raw data and the prediction result based on XGBoost is regarded as a new feature, concatenating it with the original features to form a new feature set. Then, we introduce a hybrid approach called ILSO-1DCNN, which is based on improved Lion Swarm Optimization (ILSO) and one-dimensional convolutional neural network (1DCNN). This approach is proposed for diabetes risk prediction. The ILSO-1DCNN algorithm utilizes the optimization capabilities of ILSO to automatically determine the hyperparameters of the 1DCNN network. Finally, we conducted comprehensive experiments on the PIMA dataset and compared our model with baseline models. The experimental results not only demonstrate our model's exceptional predictive performance across various evaluation criteria but also highlight its efficiency and low complexity. This study introduces a novel and effective diabetes risk prediction approach, making it a valuable tool for clinical analysis in the care of diabetic patients.

CLAHE is widely used in underwater image processing because of its excellent performance in contrast enhancement. The selection of the clip point formula is the core problem of the CLAHE methods, and the selection of suitable clipping value has become the focus of some extended methods. In this paper, an automatic CLAHE underwater image enhancement algorithm is proposed. The method determines the clipping value according to the high-order moment dynamic features of each block of the underwater image. By quantifying the dynamic features of each block in the image more precisely, and then adding it to the clipping value formula, the contrast and details of the underwater image can be effectively enhanced. In order to effectively improve the saturation and brightness of underwater images, this paper chooses a more accurate and intuitive HSV model. Experimental results show that our methods enhance the contrast subjectively, while suppressing the amplification of noise very well, and also increase the saturation of underwater images. In objective metrics, our method obtains the best values in underwater quality assessment (UIQM), SSIM, and PSNR.

Brain tumors, whether cancerous or noncancerous, can be life-threatening due to abnormal cell growth, potentially causing organ dysfunction and mortality in adults. Brain tumor segmentation (BTS) and brain tumor classification (BTC) technologies are crucial in diagnosing and treating brain tumors. They assist doctors in locating and measuring tumors and developing treatment and rehabilitation strategies. Despite their importance in the medical field, BTC and BTS remain challenging. This comprehensive review specifically analyses machine and deep learning methodologies, including convolutional neural networks (CNN), transfer learning (TL), and hybrid models for BTS and BTC. We discuss CNN architectures like U-Net++, which is known for its high segmentation accuracy in 2D and 3D medical images. Additionally, transfer learning utilises pre-trained models such as ResNet, Inception, etc., from ImageNet, fine-tuned on brain tumor-specific datasets to enhance classification performance and sensitivity despite limited medical data. Hybrid models combine deep learning techniques with machine learning, using CNN for initial segmentation and traditional classification methods, improving accuracy. We discuss commonly used benchmark datasets in brain tumors research, including the BraTS dataset and the TCIA database, and evaluate performance metrics, such as the F1-score, accuracy, sensitivity, specificity, and the dice coefficient, emphasising their significance and standard thresholds in brain tumors analysis. The review addresses current machine learning (ML) and deep learning (DL) based BTS and BTC challenges and proposes solutions such as explainable deep learning models and multi-task learning frameworks. These insights aim to guide future advancements in fostering the development of accurate and efficient tools for improved patient care in brain tumors analysis.

Chest X-ray scans are one of the most often used diagnostic tools for identifying chest diseases. However, identifying diseases in X-ray images needs experienced technicians and is frequently noted as a time-consuming process with varying levels of interpretation. In particular circumstances, disease identification through images is a challenge for human observers. Recent advances in deep learning have opened up new possibilities for using this technique to diagnose diseases. However, further implementation requires prior knowledge of strategy and appropriate architecture design. Revealing this information, will enable faster implementation and encounter potential issues produced by specific designs, especially in multilabel classification, which is challenging compared to single-label tasks. This systematic review of all the approaches published in the literature will assist researchers in developing improved methods of whole chest disease detection. The study focuses on the deep learning methods, publically accessible datasets, hyperparameters, and performance metrics employed by various researchers in classifying multilabel chest X-ray images. The findings of this study provide a complete overview of the current state of the art, highlighting significant practical aspects of the approaches studied. Distinctive results highlighting the potential enhancements and beneficial uses of deep learning in multilabel chest disease identification are presented.

Activity recognition is a fundamental concept widely embraced within the realm of healthcare. Leveraging sensor fusion techniques, particularly involving accelerometers (A), gyroscopes (G), and magnetometers (M), this technology has undergone extensive development to effectively distinguish between various activity types, improve tracking systems, and attain high classification accuracy. This research is dedicated to augmenting the effectiveness of activity recognition by investigating diverse sensor axis combinations while underscoring the advantages of this approach. In pursuit of this objective, we gathered data from two distinct sources: 20 instances of falls and 16 daily life activities, recorded through the utilization of the Motion Tracker Wireless (MTw), a commercial product. In this particular experiment, we meticulously assembled a comprehensive dataset comprising 2520 tests, leveraging the voluntary participation of 14 individuals (comprising 7 females and 7 males). Additionally, data pertaining to 7 cases of falls and 8 daily life activities were captured using a cost-effective, environment-independent Activity Tracking Device (ATD). This alternative dataset encompassed a total of 1350 tests, with the participation of 30 volunteers, equally divided between 15 females and 15 males. Within the framework of this research, we conducted meticulous comparative analyses utilizing the complete dataset, which encompassed 3870 tests in total. The findings obtained from these analyses convincingly establish the efficacy of recognizing both fall incidents and routine daily activities. This investigation underscores the potential of leveraging affordable IoT technologies to enhance the quality of everyday life and their practical utility in real-world scenarios.

Video surveillance is widely adopted across various sectors for purposes such as law enforcement, COVID-19 isolation monitoring, and analyzing crowds for potential threats like flash mobs or violence. The vast amount of data generated daily from surveillance devices holds significant potential but requires effective analysis to extract value. Detecting anomalous crowd behavior, which can lead to chaos and casualties, is particularly challenging in video surveillance due to its labor-intensive nature and susceptibility to errors. To address these challenges, this research contributes in two key areas: first, by creating a diverse and representative video dataset that accurately reflects real-world crowd dynamics across eight different categories; second, by developing a reliable framework, ‘CRAB-NET,’ for automated behavior recognition. Extensive experimentation and evaluation, using Convolutional Long Short-Term Memory networks (ConvLSTM) and Long-Term Recurrent Convolutional Networks (LRCN), validated the effectiveness of the proposed approach in accurately categorizing behaviors observed in surveillance videos. The employed models were able to achieve the accuracy score of 99.46% for celebratory crowd, 99.98% for formal crowd and 96.69% for violent crowd. The demonstrated accuracy of 97.20% for comprehensive dataset achieved by the LRCN underscores its potential to revolutionize crowd behavior analysis. It ensures safer mass gatherings and more effective security interventions. Incorporating AI-powered crowd behavior recognition like ‘CRAB-NET’ into security measures not only safeguards public gatherings but also paves the way for proactive event management and predictive safety strategies.

Accurately removing eyeglasses from facial images is crucial for improving the performance of various face-related tasks such as verification, identification, and reconstruction. This paper presents a novel approach to enhancing eyeglasses removal by integrating a mask completion technique into the existing framework. Our method focuses on improving the accuracy of eyeglasses masks, which is essential for subsequent eyeglasses and shadow removal steps. We introduce a unique dataset specifically designed for eyeglasses mask image completion. This dataset is generated by applying Top-Hat morphological operations to existing eyeglasses mask datasets, creating a collection of images containing eyeglasses masks in two states: damaged (incomplete) and complete (ground truth). A Pix2Pix image-to-image translation model is trained on this newly created dataset for the purpose of restoring incomplete eyeglass mask predictions. This restoration step significantly improves the accuracy of eyeglass frame extraction and leads to more realistic results in subsequent eyeglasses and shadow removal. Our method incorporates a post-processing step to refine the completed mask, preventing the formation of artifacts in the background or outside of the eyeglasses frame box, further enhancing the overall quality of the processed image. Experimental results on CelebA, FFHQ, and MeGlass datasets showcase the effectiveness of our method, outperforming state-of-the-art approaches in quantitative metrics (FID, KID, MOS) and qualitative evaluations.

The growing integration of the Internet of Things (IoT) in smart organizations is increasing the vulnerability of cyber threats, necessitating advanced frameworks for effective threat detection and risk assessment. Existing works provide achievable results but lack effective solutions, such as detecting Social Engineering Attacks (SEA). Using Deep Learning (DL) and Machine Learning (ML) methods whereas they are limited to validating user behaviors. Like high false positive rates, attack reoccurrence, and increases in numerous attacks. To overcome this problem, we use explainable (DL) techniques to increase cyber security in an IoT-enabled smart organization environment. This paper firstly, implements Capsule Network (CapsNet) to process employee fingerprints and blink patterns. Secondly, the Quantum Key Secure Communication Protocol (QKSCP) was also used to decrease communication channel vulnerabilities like Man In The Middle (MITM) and reply attacks. After Dual Q Network-based Asynchronous Advantage Actor-Critic algorithm DQN-A3C algorithm detects and prevents attacks. Thirdly, employed the explainable DQN-A3C model and the Siamese Inter Lingual Transformer (SILT) transformer for natural language explanations to boost social engineering security by ensuring the Artificial Intelligence (AI) model and human trustworthiness. After, we built a Hopping Intrusion Detection & Prevention System (IDS/IPS) using an explainable Harmonized Google Net (HGN) model with SHAP and SILT explanations to appropriately categorize dangerous external traffic flows. Finally, to improve global, cyberattack comprehension, we created a Federated Learning (FL)-based knowledge-sharing mechanism between Cyber Threat Repository (CTR) and cloud servers, known as global risk assessment. To evaluate the suggested approach, the new method is compared to the ones that already exist in terms of malicious traffic (65 bytes/sec), detection rate (97%), false positive rate (45%), prevention accuracy (98%), end-to-end response time (97 s), recall (96%), false negative rate (42%) and resource consumption (41). Our strategy's performance is examined using numerical analysis, and the results demonstrate that it outperforms other methods in all metrics.