Dynamic adaptive streaming over HTTP (DASH) has emerged as a foundational technology for video delivery over heterogeneous networks, particularly within 5G environments characterized by high bandwidth variability, ultralow-latency requirements, and dense mobility patterns. This survey provides a comprehensive analysis of DASH from a system-level perspective, emphasizing the interaction between adaptive bitrate (ABR) algorithms, transport-layer evolution, edge computing paradigms, and radio access dynamics. The study synthesizes existing literature across four primary dimensions: performance optimization, latency reduction, energy efficiency, and security resilience. It highlights how traditional ABR strategies—throughput-based, buffer-based, hybrid, and machine learning-driven approaches—exhibit distinct tradeoffs in quality of experience (QoE), stability, computational overhead, and power consumption. Furthermore, the evolution of HTTP protocols (HTTP/2 and HTTP/3) and the integration of multiaccess edge computing (MEC), content delivery network (CDN) caching, and device-to-device communication are shown to significantly influence streaming efficiency, albeit with new architectural complexities. A key contribution of this work is the explicit framing of DASH in 5G as a cross-layer optimization problem, where QoE, energy efficiency, and security are tightly coupled rather than independently optimizable. The survey also examines emerging security vulnerabilities across media presentation description (MPD) manipulation, segment injection, and CDN poisoning, emphasizing that DASH’s HTTP-based design inherently expands the attack surface in distributed deployments. Finally, the paper identifies persistent gaps in current research, including the dominance of simulation-based evaluations, limited real-world validation, lack of unified cross-layer frameworks, and insufficient integration of security-aware and energy-aware adaptation mechanisms. The findings suggest that future DASH systems must evolve toward coordinated, intelligence-driven architectures that jointly optimize performance, robustness, and resource efficiency in dynamic 5G and beyond networks.
With recent advancements in sensory technologies and computational power, multimodel analysis and learning have gained significant attention in the machine learning (ML) community and have been applied across a wide array of domains. Nevertheless, the effective integration of multimodal data sources poses significant challenges to the extraction of discriminative representations, as well as the design of robust fusion strategies capable of capturing complex intermodal relationships. To address these challenges, in this work, a multimodal based representation learning solution, the deep discriminative multiple canonical correlation analysis (DDMCCA), is proposed. To verify the generic naturalness and effectiveness of DDMCCA, we conduct experiments on three databases with different types of input data sources, including face recognition, handwritten digital recognition, and object recognition. Experimental results validate the superiority of the presented solution over state-of-the-art (SOTA).
This work presents a method for car license plate (LP) detection utilizing machine learning techniques that ensures rapid image processing and high detection accuracy. Three primary contributions distinguish this research. First, we introduce a novel image representation known as the “Integral Image,” which facilitates the rapid computation of features for our detector. The system is trained using a diverse set of positive (LP) and negative (non-LP) images and is validated against various real-world scenes. Second, we implement an AdaBoost-based learning algorithm to identify a minimal subset of critical visual features, leading to the development of efficient classifiers. Last, our cascade method integrates increasingly complex classifiers, enabling the rapid elimination of background regions while focusing computational efforts on promising object-like areas. The Open Computer Vision (CV) Library and Python are employed to demonstrate this implementation. We evaluated our algorithm on a dataset of 3599 images sourced from traffic video footage, achieving an impressive success rate of 95.8% in detecting Hungarian LPs.
This paper presents a field programmable gate array (FPGA)–based implementation of a deep neural network (DNN) for handwritten digit recognition. We propose the use of a fully connected four‐layer neural network with the hidden layers implementing the ReLU activation function and the output layer based on the Softmax activation function. The neural network model, including the forward propagation algorithm and the backward propagation (BP) algorithm, is entirely implemented in an FPGA system, enabling both testing and training of the network. Mathematical operations in the network are performed on 32‐bit floating‐point data. The proposed network achieves promising results in terms of system simplicity and low consumption of hardware resources while maintaining acceptable recognition accuracy and algorithm execution speed at the same time. The system is implemented on Xilinx’s ZYBO development board and achieves a precision of 92.8% digit recognition. Training and test images for system evaluation are obtained from the MNIST database. The system occupies 10,020 (57%) lookup tables, 7781 (22%) registers, 17 (21%) DSP blocks, and 3 (5%) BRAM blocks. The maximum power consumption of the system is equal to 1.602 W, at the operating frequency of the system equal to 100 MHz.
Forests uniquely deliver different vital resources, particularly oxygen and carbon dioxide purification. Wildfire is the leading cause of deforestation, where massive forest areas are annually lost due to the failure to identify and predict forest fires. Accordingly, early detection of wildfires is crucial to inform operational and firefighting teams to prevent fires from advancing. This study analyzes images taken by unmanned aerial vehicles for wildfire detection. For this purpose, the two-dimensional discrete wavelet transform was first performed on the images. Next, due to its superior ability, a convolutional neural network was utilized to extract deep features from wavelet transform sub-bands. Then, the features obtained from each sub-band were merged to create the final feature vector. Afterward, multidimensional scaling was employed to reduce the extracted non-useful features. Ultimately, the presence or absence of wildfire locations in the images was detected using proper classifiers. The proposed method reaches an accuracy and F1 score of 0.9684 and 0.9672, respectively, from the images of the FLAME dataset, indicating its efficiency in detecting the presence of wildfire locations. Thus, this method can significantly contribute to the on-time and prompt firefighting operations and prevent extensive damage to forests.
Gastrointestinal (GI) diseases are a significant global health issue, causing millions of deaths annually. This study presents a novel method for classifying GI diseases using endoscopy videos. The proposed method involves three major phases: image processing, feature extraction, and classification. The image processing phase uses wavelet transform for segmentation and an adaptive median filter for denoising. Feature extraction is conducted using a concatenated recurrent vision transformer (RVT) with two inputs. The classification phase employs an ensemble of four classifiers: support vector machines, Bayesian network, random forest, and logistic regression. The system was trained and tested on the Hyper–Kvasir dataset, the largest publicly available GI tract image dataset, achieving an accuracy of 99.13% and an area under the curve of 0.9954. These results demonstrate a significant improvement in the accuracy and performance of GI disease classification compared to traditional methods. This study highlights the potential of combining RVTs with standard machine learning techniques and wavelet transform to enhance the automated diagnosis of GI diseases. Further validation on larger datasets and different medical environments is recommended to confirm these findings.
In order to solve the problem that the current image super-resolution model has too many parameters and high computational complexity, this paper proposes a lightweight hybrid attention network (LHAN). LHAN consists of three parts: shallow feature extraction, lightweight hybrid attention block (LHAB), and upsampling module. LHAB combines multiscale self-attention and large-core attention. In order to make the network lightweight, multiscale self-attention block (MSSAB) improves the self-attention mechanism and uses windows of different sizes for group calculations. At the same time, in large-core attention, we use depth-based attention. Separate convolutions are used to reduce parameters. While keeping the receptive field unchanged, a normal convolution and a dilated convolution are used to replace the large kernel convolution. The four times super-resolution experimental results on five data sets, including Set5 and Set14, show that our proposed method performs well in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Specifically, in the benchmark data set on Urban, compared with SwinIR, the PSNR index of our method is improved by 0.10 dB. In addition, the parameter amount and calculation amount (floating point operations (FLOPs)) of our method are reduced by 315K and 16.4 G, respectively. Our proposed LHAN not only reduces the number of parameters and calculations but also achieves excellent performance in reconstruction quality.
The existing Siamese trackers express visual tracking through the cross-correlation operation between two neural networks. Although they dominated the tracking field, their adopted pattern caused two main problems. One is the adoption of the deep architecture that drives the Siamese tracker to sacrifice speed for performance, and the other is that the template is fixed to the initial features; namely, the template cannot be updated timely, making performance entirely dependent on the Siamese network’s matching ability. In this work, we propose a tracker called SiamMLG. Firstly, we adopt the lightweight ResNet-34 as the backbone to improve the proposed tracker’s speed by reducing the computational complexity, and then, to compensate for the performance loss caused by the lightweight backbone, we embed the SKNet from the attention mechanism to filter out the valueless features, and finally, we utilize the gradient-guide strategy to update the template timely. Extensive experiments on four large tracking datasets, including VOT-2016, OTB100, GOT-10k, and UAV123, confirming SiamMLG satisfactorily balance performance and efficiency, where it scores 0.515 on GOT-10k while running at 55 frames per second, which is nearly 3.6 times that of the state-of-the-art method.
Despite the extensive research on developing robust image inpainting algorithms in recent years, there are almost no objective metrics for the quality assessment of inpainted images currently. Inspired by the feature coherence in the inpainted image and the human visual perception mechanism, this paper proposes an image inpainting quality assessment (IIQA) that takes into account both visual saliency and structural features. First, the quality issues associated with image inpainting are categorized into three aspects: incoherent structure, unreasonable texture, and other results that are inconsistent with human visual perception. These quality problems are further expressed as “regions of interest” and extracted by the visual saliency method using the natural statistics model. Subsequently, the structural features are computed based on the nonlinear diffusion of the horizontal and vertical gradient field of the inpainted image. Finally, the IIQA metric incorporates brightness, gradient similarity, structural similarity, and visual saliency is established. The quality evaluation process is conducted by comparing each patch within the inpainted region with its best match from the known region. The quantitative experimental results demonstrate the effectiveness of the proposed method, especially for images with structural discontinuity. A comparative study also shows that the Spearman rank order correlation coefficient of our method achieves 0.875 on certain databases, which outperforms existing IIQA metrics.
Human abnormal action recognition is crucial for video understanding and intelligent surveillance. However, the scarcity of labeled data for abnormal human actions often hinders the development of high-performance models. Inspired by the multimodal approach, this paper proposes a novel approach that leverages text descriptions associated with abnormal human action videos. Our method exploits the correlation between the text domain and the video domain in the semantic feature space and introduces a multimodal heterogeneous transfer learning framework from the text domain to the video domain. The text of the videos is used for feature encoding and knowledge extraction, and knowledge transfer and sharing are realized in the feature space, which is used to assist in the training of the abnormal action recognition model. The proposed method reduces the reliance on labeled video data, improves the performance of the abnormal human action recognition algorithm, and outperforms the popular video-based models, particularly in scenarios with sparse data. Moreover, our framework contributes to the advancement of automatic video analysis and abnormal action recognition, providing insights for the application of multimodal methods in a broader context.
As virtual reality technology advances, 3D environment design and modeling have garnered increasing attention. Applications in networked virtual environments span urban planning, industrial design, and manufacturing, among other fields. However, existing 3D modeling methods exhibit high reconstruction error precision, limiting their practicality in many domains, particularly environmental design. To enhance 3D reconstruction accuracy, this study proposes a digital image processing technology that combines binocular camera calibration, stereo correction, and a convolutional neural network (CNN) algorithm for optimization and improvement. By employing the refined stereo-matching algorithm, a 3D reconstruction model was developed to augment 3D environment design and reconstruction accuracy while optimizing the 3D reconstruction effect. An experiment using the ShapeNet dataset demonstrated that the evaluation indices—Chamfer distance (CD), Earth mover’s distance (EMD), and intersection over union—of the model constructed in this study outperformed those of alternative methods. After incorporating the CNN module in the ablation experiment, CD and EMD increased by an average of 0.1 and 0.06, respectively. This validates that the proposed CNN module effectively enhances point cloud reconstruction accuracy. Upon adding the CNN module, the CD index and EMD index in the dataset increased by an average of 0.34 and 0.54, respectively. These results indicate that the proposed CNN module exhibits strong predictive capabilities for point cloud coordinates. Furthermore, the model demonstrates good generalization performance.
