https://doi.org/10.5573/IEIESPC.2026.15.4.467

(Jingjing Lan) ; (Peng Zhang) ; (Lingjun Wu) ; (Shuangshuang Huang) ; (Jinxi Zhang)

Aiming at the monitoring and evaluation model of water quality resources, this study deeply discusses the application of image processing transfer learning based on TOPSIS algorithm. By combining the advantages of multi-attribute decision making analysis of TOPSIS algorithm with the feature extraction ability of transfer learning, we successfully constructed an efficient and accurate water quality evaluation model. During the study, we collected 1000 water quality image data covering various water quality indicators and environmental factors for training and testing the model. Utilizing transfer learning technology, we extracted profound features from extensive natural image databases and seamlessly transferred these features to the realm of water quality image analysis. By leveraging transfer learning, we were able to successfully apply sophisticated image processing algorithms to water quality evaluation, thereby significantly enhancing the model’s accuracy and generalization capabilities. In the model evaluation, we use a variety of evaluation indicators, including accuracy, recall and F1 scores. Experimental outcomes demonstrate the efficacy of the transfer learning model, which incorporates the TOPSIS algorithm, in water quality evaluation tasks. Impressively, it achieves an accuracy rate of 92%, surpassing traditional methods by a considerable margin of 20%. Furthermore, the model exhibits remarkable improvements in recall rate and F1 score, increasing by 18% and 19% respectively. Additionally, we conducted a thorough sensitivity analysis to gain deeper insights into the model’s performance. By adjusting the weights of different water quality indexes and environmental factors, we found that the model has high stability under different water quality conditions, and has strong ability to identify slight water quality changes.

https://doi.org/10.5573/IEIESPC.2026.15.4.478

(Chunying Song) ; (Jian Liu)

In the field of apparel advertising, the accuracy of image style recognition is crucial for accurate marketing and user experience. Traditional manual feature extraction methods are inefficient and have limited accuracy. In this paper, we innovatively propose a recognition method based on improved Res-Net 152 network with migration learning. The method enhances the multi-scale feature learning capability by optimizing the network structure, and accelerates the convergence and improves the performance by using Image-Net pre-trained model. After fine-tuning the data for the girl’s clothing advertisement, the experiment shows that the recognition accuracy reaches 94.2%, which is much better than the traditional method and significantly improved than other methods (29.2%, 8.3%, 1.3%). This method not only improves the recognition accuracy, but also optimizes the training process, helping the intelligent transformation of the advertising industry, and is expected to be expanded to more advertising image recognition tasks in the future.

https://doi.org/10.5573/IEIESPC.2026.15.4.490

(Jingjing Lan) ; (Yan Huang) ; (Zhuhua Wang) ; (Mingjie Yang) ; (Jinxi Zhang)

In this study, the application of TOPSIS algorithm based on neural network in water quality resource image processing is deeply explored, and its practical application effect is evaluated through detailed data analysis. In this study, a large number of water quality resource images are used as samples, feature extraction and classification are carried out by neural network, and then TOPSIS algorithm is used for water quality assessment. A comprehensive series of experiments, coupled with rigorous data analysis, has validated the effectiveness of the TOPSIS algorithm, when integrated with neural network technology, in the realm of water quality resource image processing. The experimental findings unequivocally showcase the exceptional proficiency of this integrated methodology in the realm of processing water resource images, thereby highlighting its immense potential and broad applicability within this domain. Encompassing a comprehensive dataset of 1000 water quality image samples, which encompassed a diverse array of scenarios such as clear water, turbid water, water samples contaminated with various pollutants, and other types, the study underscores a significant leap in performance. When juxtaposed against conventional methodologies, the neural network, seamlessly fused with the TOPSIS algorithm, displays a stunning enhancement in image processing accuracy. This advancement translates into a noteworthy 15% increase, propelling the accuracy level from a solid 80% to an impressive 95%. This achievement not only validates the superiority of the proposed approach but also underscores its potential to revolutionize water quality monitoring and assessment through advanced image analysis. Simultaneously, the rate of misjudgment has been significantly reduced by 10%, falling from 15% to just 5%. Notably, in the recognition of clear water samples, the algorithm achieves an astonishing accuracy rate of 98%, while turbid water samples are recognized with a 95% accuracy. Even for water samples containing various pollutants, the average accuracy rate remains impressive at 92%. Furthermore, this algorithm demonstrates superior processing speed, enhancing it by nearly 30% compared to traditional methods.

https://doi.org/10.5573/IEIESPC.2026.15.4.503

(Zhiqiang Wu) ; (Teng Gong) ; (Yunfan Chen)

3D object detection is one of the key techniques in autonomous driving. However, existing 3D object detectors will likely be impaired under low illumination conditions and show poor performance on small object detection. To resolve this limitation, this paper proposes a tone mapping-based fine disparity R-CNN (TFD R-CNN) for effective 3D object detection. We first develop a tone mapping technique to enhance the color contrasts and thus to make objects clearer, for easier detection of low illuminated objects. Furthermore, optimization of grouping radius has been done to improve the detection performance on small-sized objects. By reducing the grouping radius, the features of small objects can be preserved, so that more refined feature information can be extracted. Experimental result comparisons show that the proposed TFD R-CNN outperforms the well-known Disp R-CNN. Specifically, the proposed method can achieve about 4% better performance in average precision than the baseline Disp R-CNN on the KITTI dataset.

https://doi.org/10.5573/IEIESPC.2026.15.4.512

(Youngshin Kang) ; (Cheolsoo Park)

The estimation of blood pressure using pulse transit time is critical for the continuous monitoring of blood pressure during our daily lives. This study addresses the challenge by introducing a novel, non-invasive approach for blood pressure estimation that does not rely on pulse transit time (PTT). Instead, we propose an innovative graph-based neural network architecture that leverages the interconnectedness of multiple physiological signals, specifically ballistocardiogram, photoplethysmogram, and electrocardiogram. Specifically, the adjacency matrix for the graph neural networks is constructed with one-way and two-way directional relationships among the physiological signals, Euclidean distance, and causality. The findings suggest that our graph-based neural network model holds significant potential for enhancing continuous, non-invasive blood pressure monitoring, thereby contributing to better cardiovascular health management in everyday life.

https://doi.org/10.5573/IEIESPC.2026.15.4.518

(Guang Zhang) ; (Yukun Yang)

This study enhances the efficiency and accuracy of English automatic translation systems by integrating multiple interactive features with a natural language processing (NLP) algorithm. As globalization intensifies, the demand for effective cross-language communication increases, making translation system optimization crucial. Despite advancements in AI translation, challenges remain in efficiency, semantic understanding, and context mastery. This paper addresses issues such as efficiency lags, semantic misunderstandings, and context constraints, particularly when navigating cultural nuances?like the different meanings of "dragon" in Chinese and Western cultures? and idioms, such as the literal versus Italian translation of "break the ice." Traditional methods often struggle to convey the original intent accurately. Moreover, failures in context perception, especially in complex or implied contexts, lead to inaccuracies. To overcome these challenges, we propose an NLP algorithm that dynamically integrates syntactic, semantic, and pragmatic features, thereby improving translation accuracy and efficiency. This study offers valuable insights and approaches for optimizing translation systems and advancing natural language processing research.

https://doi.org/10.5573/IEIESPC.2026.15.4.532

(So-Yeon Jang) ; (Jong-Ok Kim)

Recent vision-language models are increasingly applied to dense prediction tasks such as semantic segmentation, where textual input provides high-level semantic guidance. A common strategy is to use cross-attention mechanisms to integrate image and text features. However, such methods often suffer from limited spatial adaptability and substantial computational overhead. To resolve these issues, we propose an architecture composed of a Wavelet-Aware Context Construction (WACC) module and a Cross Deformable Fusion (CDF) module. WACC generates enriched textual representations by decomposing visual features into multiple frequency components via Discrete Wavelet Transform (DWT). It captures both global semantic layout and fine-grained structural cues. These informative text features are subsequently fused with image features using CDF, which employs deformable convolution to achieve spatially adaptive and content-aware alignment across modalities. This design enables more precise and efficient cross-modal interaction, leading to improved segmentation performance. Experimental results demonstrate that our method achieves an mIoU of 0.790, exceeding the baseline of 0.777 and confirming the effectiveness of the WACC and CDF modules.

https://doi.org/10.5573/IEIESPC.2026.15.4.541

(Rui Zhang) ; (Dongkai Liu) ; (Jingsha Zheng) ; (Yan Wang) ; (Gang Li) ; (Jinchuan Huang) ; (Xulingyun An) ; (Gengliang Li)

Recent advancements in the Internet of Things (IoT) have opened new possibilities for digital agriculture, significantly impacting the animal husbandry economy and grassland ecosystem stability. Traditional grassland evaluation methods are often time-consuming, labor-intensive, and prone to inaccuracies. To address these challenges, we propose a novel framework that employs a fuzzy data fusion approach for grassland evaluation. This method considers multiple factors?temperature, humidity, light, air pressure, and the Normalized Difference Vegetation Index (NDVI)?with weights determined through the Analytic Hierarchy Process (AHP). Our framework integrates an environmental monitoring system that leverages NB-IoT, sensor technology, and image processing techniques. Experimental results demonstrate that our method captures more comprehensive grassland information and overcomes the limitations of single-factor assessments, thereby enhancing the reliability and accuracy of grassland grade evaluations. By providing real-time, precise, and holistic assessments, the integration of AHP and fuzzy data fusion significantly improves the overall reliability of grassland evaluations.

https://doi.org/10.5573/IEIESPC.2026.15.4.554

(Chuanhui Qiu)

Tourism development on the Mashan Peninsula has led to significant landscape changes, with increased construction land encroaching on other types, especially broad-leaved forests. This has resulted in a fragmented, heterogeneous landscape with reduced connectivity and dominance. These changes are driven by tourism construction, agricultural demand, and policy promotion, which have also led to a decline in ecosystem service value (ESV), particularly in regulation services. Correlation and regression analyses help identify key landscape indices affecting ESV, offering strategies to optimize landscape patterns and improve ecosystem service efficiency. For tourism-oriented rural landscapes, constructing “ecological safety patterns” and “cultural landscape safety patterns” is essential. The former ensures ecosystem stability, while the latter protects cultural resources. After applying contrastive learning, the landscape diversity index increased from 0.65 to 0.75, and ecological corridor connectivity improved from 60% to 80%. Agricultural productivity rose by 10%, with yields increasing from 2.5 to 2.75 tons per hectare, and water resource efficiency improved from 70% to 82%. By overlaying ecological and cultural safety patterns, comprehensive zoning strategies can promote balanced ecological protection and cultural development, offering valuable insights for sustainable rural landscape planning.

https://doi.org/10.5573/IEIESPC.2026.15.4.567

(Somi Lim) ; (Jihyeon Ryu)

Timeline reconstruction is fundamental to digital forensic investigations, as timestamps are critical for establishing the chronological sequence of events. However, several malware and adversaries exploit inherent deficiencies of this procedure by employing timestamp manipulation techniques to conceal malicious activities and mislead forensic examinations. Therefore, research into timestamp forgery is essential. While existing literature focuses on individual tools or detection methods, there is a lack of comprehensive research integrating these perspectives. This study investigates forgery techniques and tools? such as XTST, SetMACE, and Timestomp?incorporating those utilized by prominent threat actors such as APT29 and the Lazarus Group. Furthermore, we classify detection methodologies and evaluate them across four practical scenarios. By analyzing their respective strengths and limitations, this paper provides a strategic roadmap for future research in timestamp forensics.

https://doi.org/10.5573/IEIESPC.2026.15.4.582

(Seongmo Park) ; (Piljae Park) ; (Kyung-Hwan Park)

In this paper, we propose a novel hardware architecture specifically designed to enable real-time response in disaster environments, where rapid decision-making and low-latency processing are critical. Our approach focuses on optimizing a GAN-based radar voice imaging system, which processes radar signals and generates visual representations in real time. To enhance accuracy and robustness, we developed a simulation framework that assigns dynamic weights to each incoming signal, enabling the system to identify the signal path with the minimum loss through iterative optimization. The proposed architecture was implemented using Register-Transfer Level (RTL) coding, targeting efficient hardware utilization while maintaining high-speed operation. To validate the design, we conducted logic synthesis and timing verification using the Xilinx VivadoR synthesis environment. This step ensured that the proposed system not only met functional correctness but also satisfied timing and area constraints required for deployment in edge computing scenarios, such as portable disaster response units. The synthesis results demonstrated that the design is highly resource-efficient, consuming only 2% of available Look-Up Tables (LUTs), 2% of flip-flops (FFs), 13% of input/output (I/O) resources, and 1% of global clock buffers (BUFGs). These results indicate the architecture’s suitability for integration into FPGA-based platforms with limited hardware resources, offering a practical solution for real-time radar signal processing in critical field applications.

https://doi.org/10.5573/IEIESPC.2026.15.4.591

(Tuan Vo) ; (Hyouk-Kyu Cha)

This paper presents the design of a bipolar current-mode stimulator integrated circuit (IC) for implantable prosthetic devices using 0.18-μm standard CMOS process. The proposed stimulator IC is capable of delivering up to 1-mA current through 10-kΩ load using 12.8 V supply voltage. The output impedance of the stimulator is enhanced by utilizing a regulated cascode output current driver stage, which achieves a voltage compliance of 12.4 V. In order to allow low-voltage standard transistors to operate reliably at high voltage, transistor stacking and dynamic gate biasing techniques are used. In addition, a simple active charge balancing circuit is included to achieve minimal net charge after stimulation cycle is complete. The overall layout area is 0.12 mm2 .