https://doi.org/10.5573/IEIESPC.2025.14.5.569

(Hongwu Hou)

In recent years, as the Internet technology develops, image processing software has gradually become popular, which provides convenience for plagiarism and tampering of images, and also brings difficulties to the protection of intellectual property rights. Therefore, how to quickly and accurately detect plagiarized and tampered images has become a top priority. In order to achieve accurate detection and location of plagiarized images, this paper puts forwards an image plagiarism detection model with KAZE and polar harmonic transform. Experimental results show that in the CoMoFoD database, the KAZE-polar harmonic transform algorithm can accurately identify images modified by copy-paste, brightness tampering, and blur processing. Compared with fuzzy C-means clusteringEmperor Penguin optimization-block feature matching algorithm and sparse recovery-key point matching algorithm, the average accuracy of KAZE-polar harmonic transform algorithm is 98.1%, which is higher than other algorithms, and each image The image recognition time only takes 10.86 s, which is faster than other algorithms. In the Coverage database, the KAZE-polar harmonic transformation algorithm can accurately identify and locate images with background tampering, and can effectively resist the interference of real similar objects; in addition, its average accuracy and image recognition speed are 90.3% and 11.5 s / Zhang, outperforms other algorithms. The above results show that the KAZE-polar harmonic transform algorithm proposed in the study can quickly and accurately identify and locate tampered pictures and tampered parts.

https://doi.org/10.5573/IEIESPC.2025.14.5.579

(Qun Luo) ; (Zhendong Liu)

With the continuous growth of medical data, traditional learning methods have limitations in processing private data and model training. A federated learning differential privacy medical image classification algorithm based on hybrid reality technology was proposed to conduct medical image analysis in a distributed environment. Firstly, the differential privacy medical image classification based on federated learning was studied. Then, the results of the differential privacy medical image classification model based on federated learning were analyzed. The results confirmed that as the learning rate increased, the accuracy of the research algorithm gradually increased. When the learning rate was 0.1 and the iterations were 10, the research algorithm reached the highest accuracy, which was 99.82%. In addition, the study also analyzed the impact of the number of clients on algorithm performance. In the range of 8 to 12 clients, this research algorithm demonstrated significant advantages compared to traditional image classification algorithms. When the number of clients was 8, its AUC value reached its highest, at 92.23%. In summary, the research algorithm has shown significant advantages in medical image classification tasks.

https://doi.org/10.5573/IEIESPC.2025.14.5.592

(Aditi Gupta) ; (Hoor Fatima)

Ovarian cysts pose a serious health risk to women of all ages and backgrounds all around the globe. In medical diagnosis, distinction of ovaries affected by cysts from the normal ovaries through ultrasound images is a crucial task. Nowadays, deep learning models have emerged as effective tools for classification tasks in the field of medical imaging. This paper employs Residual Network (ResNet), EfficientNet, Densely connected Network (DenseNet) and Visual Geometry Group (VGG) models that have depicted impressive performance in automating the classification process. The efficacy of the models in detecting the existence of ovarian cysts is assessed by the performance standards, namely accuracy, precision, recall, specificity, f1-score and area under receiver operating curve. DenseNet-169 has proved to be the best deep learning model for precise and efficient classification of ovarian cysts with accuracy 99.78% in ultrasound images. This research contributes a major milestone in the classification of cystic ovaries from normal ovaries and henceforth, leads in the advancement of diagnostic capabilities in the field of ovarian pathology.

https://doi.org/10.5573/IEIESPC.2025.14.5.603

(Yige Wang)

This study introduces a Hybrid Deep Convolutional Neural Network (HDCNN) for advanced vibration signal analysis and fault detection in mechanical systems. HDCNN combines One Dimensional Convolutional Neural Network (1DCNN), Two-Dimensional Convolutional Neural Network (2DCNN), and Deep Neural Networks (DNN) to enhance fault diagnostic efficiency. The 1DCNN is used for regression tasks like analyzing machining surface roughness, while 2DCNN handles classification tasks such as tool wear and bearing fault diagnosis using timefrequency images from short time Fourier transform (STFT). DNN integrates features from 1DCNN and 2DCNN for effective regression and fault classification. Experimental results show that HDCNN outperforms existing techniques like CNN, DNN, FANN, PSONN, and GANN in diagnostic accuracy, convergence speed, and stability. This research highlights HDCNN as a powerful tool for predictive maintenance and real-time fault identification in mechanical systems, showcasing the advantages of deep learning in mechanical fault diagnostics.

https://doi.org/10.5573/IEIESPC.2025.14.5.616

(Huimin Chang) ; (Qihui Ding)

Face Super-Resolution (FSR) methods based on deep learning have made significant progress in recovering severely degraded facial images. However, existing approaches still face challenges when dealing with extremely low-resolution and noisy inputs, particularly in preserving facial structures and fine details. This paper introduces a Restormer-based Face Super-Resolution (RFSR) method that integrates the robust feature extraction capabilities of Transformers with the temporal processing advantages of recurrent neural networks. The RFSR architecture comprises four key components: an initial feature extraction module (G1), a Restormer module, a Recurrent SuperResolution module (RecurrentSRModule), and a final reconstruction module (G2). The Restormer module employs multi-head transposed self-attention mechanisms to capture long-range dependencies, effectively extracting global facial features. The RecurrentSRModule refines and enhances image details through multiple iterations. This iterative collaboration mechanism enables the network to improve reconstruction quality progressively, particularly when processing challenging low-quality inputs. Additionally, the network incorporates a residual connection that adds the upsampled original input to the network output. This design allows the main network to focus on learning highfrequency details and image enhancement while preserving low-frequency information from the original input. This approach improves reconstruction stability. It also enhances detail fidelity in the super-resolved images. Extensive quantitative and qualitative experimental results demonstrate that the proposed RFSR method outperforms existing state-of-the-art FSR approaches in recovering high-quality facial images, especially when processing extremely low-resolution and heavily noisy inputs. Our method effectively restores facial structures and texture details while maintaining identity consistency and subtle expression variations.ㅍ

https://doi.org/10.5573/IEIESPC.2025.14.5.631

(Ying Liu) ; (Lin Liu)

Using remote sensing technology to obtain and identify plant images can enhance data for landscape design and streamline the design process. To address the issue of low recognition accuracy in plant remote sensing models, this study developed an adaptive threshold binary mask algorithm using mixed maximum inter-class variance and an image classification algorithm based on an improved ResNet50 network, combining them into a plant remote sensing image recognition model. The test results showed that the average recognition recall and F1 mean of the designed model on various plant remote sensing images in the test set were 97.8% and 97.7%, respectively, which were 4.2% and 4.1% higher than the method ranked second in overall numerical values. The offline area of the receiver operation characteristic curve of the designed model on the test set was 73.5%, which was 8.4% higher than the algorithm before improvement. From the test results, the recognition model designed this time has stronger recognition ability than the currently commonly used models. This model can be used to assist landscape design, helping designers identify the distribution of plants in the landscape location and the survival status of green plants after construction.

https://doi.org/10.5573/IEIESPC.2025.14.5.644

(Doohong Park) ; (Donggoo Kang) ; (Joonki Pai)

This paper explores the influence of South Korea’s defense reform initiative, “Defense Innovation 4.0,” by analyzing media coverage before and after its establishment as a national agenda. We collected articles related to defense from a year prior to a year following the announcement. Utilizing the KeyBERT model, we extracted key terms to guide our application of Latent Dirichlet Allocation (LDA) for topic modeling. This approach allowed us to discern the evolving discourse on defense issues, particularly in relation to “Defense Innovation 4.0.” Furthermore, we employed a BERT-based model to assess the content of the “Defense Innovation 4.0” policy against the backdrop of defense discussions pre- and post-announcement. Our findings reveal a notable surge in media attention towards the defense sector, with a significant shift in the thematic focus of private media coverage aligning with the policy’s introduction.

https://doi.org/10.5573/IEIESPC.2025.14.5.657

(Xuyun Gon)

Given the existing problems such as weak interaction ability, low information recognition rate, and poor accuracy of goods identification, a new system of goods identification based on a multi-modal learning method was designed. By integrating commodity visual, voice, text, and other multiple information, a multi-modal information recognition model is built, commodity data is mapped to the model analysis module using modal retrieval, ITC contrast training, ITM interactive training, and ITG weighted training is carried out, and multi-modal commodity information is initially integrated. SDI+HDMI dual-interface encoder was used to encode the commodity graphic frequency information, and TRIOPC-MCAT-2 controller and IC identifier were selected to optimize the hardware equipment, effectively improve the commodity identification computing power and information response rate, and enhance the stability of the system. The multimodal learning model is used to transform the commodity information into 10-dimensional mapping vectors, multimodal coding is carried out according to the model input hierarchy, key features are extracted using SVM classifier, and multimodal feature information fusion vectors of commodities are obtained through interactive guided weighting operation, and commodity recognition is carried out according to the training results. According to the experimental results, given the massive commodity information of multiple categories, the success rate of the multi-modal learning-based live broadcast cargoes identification system designed in this paper for the commodity multi-modal information fusion recognition has reached more than 88%, the recognition time is less than 90ms, and the recognition accuracy rate is higher than 95%, indicating that the system studied in this paper has good recognition performance. The practical application effect is better than the traditional method, which can meet the current needs of live broadcast goods identification.

https://doi.org/10.5573/IEIESPC.2025.14.5.668

(Adeel Munawar) ; (Mongkut Piantanakulchai) ; (Muhammad Nadeem Ali) ; (Byung-Seo Kim)

The transport system is the key element to the economy of any country and requires significant attention in various aspects, like road safety, affordability to the common person, sustainability, fuel consumption, climate, pollution, and many more. Shared mobility is one of the promising solutions that can influence many aspects and is formally known as carpooling. Carpooling is a promising solution for reducing urban transportation energy use, producing less pollution, and a smaller number of vehicles (easy to manage the infrastructure). Thanks to information communication and technology and smart mobile devices that enabled ?smart carpooling? leverage this technology to more efficient, and widespread use. Smart carpooling has strong potential to reduce energy use and environmental impacts from private urban transportation, however, requires critical investigation in various aspects. To effectively implement smart carpooling, it is imperative to delve into the theoretical underpinnings, assess its advantages and disadvantages, draw insights from real-world case studies, and identify the challenges and obstacles associated with its deployment. This review paper aims to assess smart carpooling research areas, uncover new research directions, challenges, and solutions, and provide insights to the research community

https://doi.org/10.5573/IEIESPC.2025.14.5.679

(Li He)

This article explores web service recommendation algorithms for heterogeneous environments by incorporating a time dimension into datasets, transforming static data into dynamic elements. This approach improves both the accuracy of service quality predictions and the speed of service recommendations, enhancing overall efficiency. The study introduces the DBSLF algorithm, which integrates cross-platform data features while prioritizing privacy protection. By fusing user service data from diverse edge server platforms, the algorithm provides more precise recommendations through multi-platform data analysis. The algorithm’s effectiveness is validated through experiments with varying data densities, significantly reducing the mean absolute error (MAE) for missing Quality of Service (QoS) values. To address service recommendation challenges in single-platform heterogeneous environments, the concept of continuous time slices is introduced, utilizing historical data to predict missing QoS values. Additionally, a privacy-preserving, cross-platform recommendation algorithm is proposed, leveraging latent factor models, density-based clustering, and matrix decomposition techniques. This approach enhances prediction accuracy, mitigates cold start issues, and contributes to smart transportation networks by improving road congestion predictions and enhancing daily travel experiences.

https://doi.org/10.5573/IEIESPC.2025.14.5.692

(Tingting Zhao)

This paper introduces an innovative urban traffic congestion prediction model, which combines knowledge graph and graph neural network technology to improve prediction accuracy and traffic management efficiency. Firstly, the data preprocessing steps are described in detail, including the integration and standardization of traffic, weather and event data. Then, the traffic knowledge graph is designed, the entity, relationship and ontology structure are defined, and the dynamic update mechanism is constructed. The core of the model adopts GNN combined with LSTM, embedding entity relations through TransE, learning spatiotemporal features through graph convolutional network, and capturing time series dynamics through RNN. The experimental design is based on a one-year comprehensive traffic dataset. The results show that the proposed KG-GNN model performs well in prediction accuracy, stability and interpretation, significantly reduces the average congestion time and delay index, and improves public satisfaction. Case studies further demonstrate the effectiveness of the model in easing congestion, optimizing public transport and enhancing travel experience.

https://doi.org/10.5573/IEIESPC.2025.14.5.705

(Yein Choi) ; (Sung Soo Hwang)

This paper explains recent researches that integrate multiple sensors in a field of simultaneous localization and mapping (SLAM), which is of high interest in areas such as autonomous driving. Basic sensors commonly used in SLAM, such as LiDAR, camera, and inertial measurement unit (IMU), have individual drawbacks when used as single sensors. Therefore, in many SLAM researches, these shortcomings are overcome by fusing different sensors to complement each other and enhance performance, aiming for more accurate state estimation. Various methods are available for optimizing the information from each sensor during this process. In this paper, we aim to explain methods for integrating sensors information such as MAP, Kalman Filter, and MLE. Moreover, we will introduce research that utilizes information obtained from sensors. We hope that this paper seeks to understand the types of sensor data fusion methods employed when multiple sensors information is available.