https://doi.org/10.5573/IEIESPC.2026.15.5.599

(Yuhao Lv) ; (Hancheng Yu) ; (Mingkui Feng) ; (Ming Zhou)

Ellipse detection holds significant importance in image processing and plays a key role across various application scenarios. However, achieving accurate and robust ellipse detection remains challenging. Traditional methods can perform precise detection in controlled environments, but they often struggle with ellipses in complex scenes. Conversely, existing deep learning-based approaches demonstrate improved robustness in handling complex scenarios; however, they frequently exhibit limitations in detection accuracy. To address these challenges, this paper presents a novel auxiliary branch based on edge extraction to enhance feature learning and parameter regression for ellipses, ultimately achieving high-precision detection. The auxiliary branch consists of two fundamental components: an edge extraction module and a rotated region of interest (RoI) parameter regression module. The edge extraction module leverages a U-Net architecture to effectively capture and refine elliptical edge features, enabling precise learning of target contours. Simultaneously, the rotated RoI parameter regression module integrates these edge features with preliminary detection results from the first stage to perform accurate parameter regression, thereby improving localization performance. Experimental results demonstrate that the proposed method significantly enhances detection performance, particularly under high overlap ratios.

https://doi.org/10.5573/IEIESPC.2026.15.5.609

(Woo-Seok Lee) ; (Maaz Ahmed) ; (Kang-Sun Choi)

This paper introduces a novel loss function, the smoothed boundary distance loss, designed to enhance character segmentation in high-resolution scans of old printed documents. Our method calculates deviation between the boundary of predicted character strokes and the counterpart of the ground truth strokes. Diverging from traditional Euclidean distance metrics, our approach determines the deviation indirectly utilizing boundary pixel value difference over a Gaussian smoothed version of the ground-truth for computational efficiency. The proposed method not only smoothens the boundary lines in character extraction but also effectively eliminates noise and outliers, significantly improving the clarity and accuracy of the segmentation process

https://doi.org/10.5573/IEIESPC.2026.15.5.618

(Shaila Sharmin) ; (Mohammad Mohiuddin) ; (Md Saiful Islam) ; (Jia Uddin)

Since brain tumors have a variety of forms, it is crucial to determine the types of tumors. The convolutional neural network (CNN) can extract features automatically with high-performance accuracy. The main problem of CNN is that it is unable to extract unknown features both locally and globally, with a large number of parameters. As a result, it must deal with computational complexity and take a long time to train, and additionally, due to a large number of parameters, which can lead to overfitting. To address this, a novel lightweight Deep Parallel CNN is designed to extract the known features both locally and globally and introduce Global average pooling technology instead of the fully connected layer, which effectively reduces the number of parameters and solves the overfitting problem. In the initial stage, resized data (60×60) is augmented to enhance the number of images in the pre-processing step. Next, the augmented data is applied to the DPCNN to extract the features both locally and globally in the feature extraction step. Lastly, different forms of classifiers like SoftMax, KNN, and SVM are used to classify the various forms of tumors from MRI scans in the classification step. The proposed DPCNN-SVM architecture achieved a classification accuracy of 97.7% which is compared with other existing models.

https://doi.org/10.5573/IEIESPC.2026.15.5.629

(Xiang Feng) ; (Maoyi Li)

To effectively reduce the storage space of 3D animation and improve transmission speed, this study proposes a 3D animation image optimization and compression method that combines the e K-means algorithm and dual optimization compression algorithm. The e-K-means algorithm optimizes the structure of 3D animated images by ensuring that each cluster is of equal size. The dual optimization compression algorithm combines lossless compression and lossy compression techniques. The results indicated that under certain conditions, the research method had reconstruction errors of 0.26% and 0.41% when processing 3D animated images of characters and landscapes. In the actual processing of 3D animation, the average compression rate reached 47.83%, CPU usage was less than 38%, the average video quality evaluation index reached 92.06, and the average processing time was 14.53 minutes. In summary, the research method can achieve good compression results with lower resource consumption and shorter compression time while ensuring the quality of 3D animation images. This study provides a new technology for efficient compression of 3D animation images, achieving a good balance between quality, efficiency, and resource utilization. It also offers novel solutions for the applications in related fields such as file compression and video processing.

https://doi.org/10.5573/IEIESPC.2026.15.5.642

(Xiaoying Cai)

This study explores the implementation of a green supply chain management system using a hyperparameter-optimized neural network model. By integrating neural network algorithms and optimization techniques, the study enhances prediction accuracy and efficiency. An in-depth analysis identifies key green supply chain components and addresses shortcomings in existing frameworks through a composite model architecture. This study meticulously constructs a model capable of navigating intricate supply chain data by incorporating a diverse array of neural network algorithms, notably Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs). Subsequently, to refine the model’s parameters and achieve peak performance, sophisticated hyperparameter optimization techniques, including mesh search and Bayesian optimization, are employed. This comprehensive approach ensures that the model is tailored to excel in managing the complexities inherent in supply chain data analysis. The green supply chain management system, enhanced by a hyperparameter-optimized neural network model, demonstrates high predictive accuracy and resilience. It effectively forecasts environmental risks, energy consumption, and resource use while offering decision support for managers. This study introduces a new technical approach, strongly supporting sustainable enterprise development and green supply chain transformation.

https://doi.org/10.5573/IEIESPC.2026.15.5.656

(Liping Deng)

The English education platform model recommendation algorithm stands out as a pivotal tool in the realm of Internet information services owing to its remarkable adaptability within the realm of big data. Through a meticulous examination of various recommendation algorithms, including the mainstream ones, and specifically targeting the English education platform model algorithm, groundbreaking advancements have been achieved. Amidst intricate challenges of training in vast data environments due to an excess of parameters, an innovative approach has been proposed. Enter the FGBLFM algorithm?a refined version based on function space gradient enhancement. This algorithm outshines traditional models by offering superior insights into user-item relationships and enhancing the accuracy of inference ratings within English education platforms. Noteworthy improvements have been made in spelling error detection through the integration of named entity recognition. Identification accuracy has surged, reducing errors in recognizing proper nouns by a considerable margin. Furthermore, error correction suggestions saw a significant accuracy boost, showcasing enhancements from 0.862 to 0.941?a substantial 9.16% rise in accuracy. Delving deeper, the FGBLFM algorithm has been successfully implemented on the Spark distributed processing platform. By utilizing the Movie Lens dataset for evaluation with RMSE and MAE metrics, comparisons against collaborative filtering methods and existing English education platform models unveiled the algorithm’s effectiveness. Through this holistic approach, the FGBLFM algorithm has solidified its position as a game-changer in the landscape of recommendation algorithms, marking a paradigm shift in driving heightened accuracy, enhanced performance, and usability within English education platforms.

https://doi.org/10.5573/IEIESPC.2026.15.5.670

(Jun Li)

Aiming at the insufficient visualization effect of basketball game data, this paper proposes a mixed prediction model based on voting ensemble learning and improved random forest. Combined with voting ensemble algorithm to process data, the final result of basketball game data is obtained. The experiment showed that the average accuracy of the model reached 0.856, the average absolute error was stable within 0.021-0.041, and the response time of real-time event data visualization was less than 0.8 seconds. The core innovation of the research method is to propose a cluster optimization algorithm to solve the inefficient classification of unlabeled data in traditional random forests. A mixed voting strategy is designed, where hard voting is used for discrete event classification and soft voting is used for continuous index prediction to enhance the model generalization ability of complex event scenarios.

https://doi.org/10.5573/IEIESPC.2026.15.5.683

(Wenyan Zhao) ; (Xiaowen Zhang)

Currently, research on intelligent educational content recommendation systems suffers from issues like inadequate content comprehension and inaccurate personalized recommendations, primarily due to the lack of effective content understanding and user behavior analysis techniques. This study innovatively integrates deep learning and natural language processing to address these challenges. Deep learning models enhance text content analysis, while natural language processing enables precise exploration of user needs. Leveraging a vast dataset encompassing students’ online learning trajectories and educational text information, the proposed system combines deep neural networks with natural language processing to dissect text semantics and structures, identify learners’ interests and knowledge deficiencies, and rapidly retrieve optimal learning materials from millions of educational resources. Empirical results demonstrate significant improvements: the AUC value, which gauges the system’s ability to differentiate suitable from unsuitable resources, increases by 23%, and the click-through rate of recommendation lists rises by approximately 19% compared to traditional collaborative filtering methods. The system-generated recommendations better align with users’ actual requirements, boosting learning efficiency and satisfaction, and enhancing the accuracy and personalization of intelligent educational content recommendation systems

https://doi.org/10.5573/IEIESPC.2026.15.5.701

(Fazeela Mughal) ; (Ghulam Musa Raza) ; (Muhammad Waseem Raza) ; (Byung-Seo Kim)

Machine learning techniques enhance security in Software-Defined Networks by detecting attacks. Among various network threats, Distributed Denial of Service (DDoS) attacks are the most prevalent, causing monetary loss and server outages. Extensive research has focused on detecting these attacks using real-time and synthetic datasets. This paper evaluates the performance of machine learning algorithms in detecting DDoS attacks with high accuracy. A simulated dataset was generated using Packet Sender, and five benchmark datasets were selected: three real-time datasets (CICDDOS2019, BoT-IoT, and ToN-IoT) and two synthetic datasets (Mendeley and Bogazici University DDoS Dataset). Seven machine learning algorithms?KNN, Decision Tree, Random Forest, Logistic Regression, MPL, XGBoost, and SVM?were trained and tested. Results are based on accuracy, considering the best features within each dataset instead of direct feature comparison across datasets. Since each dataset has distinct features, using identical features for all datasets may yield biased results. Some features may be highly relevant in one dataset but less significant in another. This study highlights the importance of selecting optimal dataset-specific features to improve attack detection accuracy while mitigating bias in comparative analysis.

https://doi.org/10.5573/IEIESPC.2026.15.5.712

(Wei Zhang) ; (Jinyong Ou) ; (Kaifeng Deng)

At present, with the rapid development of social economy, the demand for electricity is increasing, in order to ensure the stable operation of the power system, at present mainly rely on inspection personnel to maintain the equipment. Currently, the low level of information ionization and the cumbersome process hinder the timely transmission of inspection data. This results in a lack of on-site operational records, making it difficult to trace back problems. Additionally, the operation lacks sufficient supervision, thereby compromising the efficiency and operational safety of personnel staff. In this paper, a power inspection personnel detection system based on BeiDou GNSS positioning technology and augmented reality technology is constructed in response to the above problems, in order to monitor and command the whole process of power operation, improve the efficiency of power operation and reduce the probability of power accidents. The system utilizes BeiDou GNSS to realize positioning of power inspectors and assists inspection through augmented reality technology, and then utilizes mobile intelligent terminal to record and broadcast the whole process of power operation to realize monitoring and commanding of power operation.

https://doi.org/10.5573/IEIESPC.2026.15.5.727

(Yan Zhao) ; (Jianghao Yang) ; (Xiaofeng Chen)

The wake effect of wind turbines in wind farms has a significant impact on wind energy utilization. Existing optimization methods suffer from low efficiency and large errors in handling coordinated control of multiple units. To this end, a mathematical model for the superposition effect of wind turbine wake is established. Then a multi-unit coordinated optimization active power control method is proposed by combining K-means clustering and improved particle swarm optimization algorithm to address this challenge. The results showed that the wind speeds of the six wind turbines in the wind farm exhibited a downward trend under the influence of wake flow. When the wind turbine operated at reduced load and reduced the wind energy conversion, the wind energy utilization coefficient decreased. The active power of the improved model gradually increased at 1.22s-1.60s, and remained consistent with the true value. The relative error was 0.41%, while the relative error of the unimproved model was 2.26%. After 1.628s, the active power curve of the improved model remained consistent with the true value. The relative errors between the improved model, K-means model, and the true values were 0.41% and 1.82%, respectively. This method can flexibly switch the active power target of the wind farm, optimize the unit speed and pitch angle, improve the system frequency response, and increase the output power of the wind farm. This achieves optimized computational efficiency and accuracy, reduces reliance on traditional energy sources, and promotes the development of sustainable energy. Applying this method to actual wind farms can ensure its feasibility and effectiveness in practical engineering, promoting the development and application of wind power generation technology.

https://doi.org/10.5573/IEIESPC.2026.15.5.741

(Jin Peng) ; (Yang Zhou) ; (Yangyi Zhang)

This research presents a new Maximum Power Point Tracking (MPPT) algorithm for photovoltaic systems under partial shading, which is a hybrid of Genetic Algorithm (GA), Grey Wolf Optimization (GWO), and Incremental Conductance (INC). To improve global diversification, the GA-GWO-INC MPPT uses GA-based crossover and mutation. It also utilizes GWO for quicker convergence and INC for accurate local optimization. Different partial shading patterns were used to test the algorithm, and the results obtained were also used to compare the performance of the proposed method with those of traditional INC, GWO, and GA-GWO-INC based MPPT methods. The outcomes indicate that the proposed GA-GWO-INC MPPT attains a tracking accuracy of up to 99.998%, outperforms other methods in all tests, and delivers a faster response with a computing time reduced by 9.4%-34%. In addition, the method keeps away from local optima while still achieving high accuracy and thus is the best performer in dynamic partial shading situations. The main discoveries demonstrate that GA-GWO-INC MPPT strategies have the potential to raise the efficiency, stability, and reliability of photovoltaic power systems to a great extent under complex shading conditions.