https://doi.org/10.5573/IEIESPC.2025.14.4.431

(Zemei Liu)

The intelligent design of logos has great application prospects in product analysis and product authenticity identification. To achieve more effective logo design, the Discriminatory Regional Guidance and Enhancement Network (DRGE-Net) was introduced for feature extraction and classification of images, and a Logo YOLO detection method was proposed for precise detection of image targets. It was proved that in the performance experiment of validating the logo classification method, the classification accuracy of DRGE-Net in the four best performing logo categories was above 80%, and the highest reached 92.68%. At the same time, among the four worst performing logo categories, the classification accuracy of DRGE-Net was higher than 50%, and the highest reached 60.8%. In the experiment to verify the function of the logo detection method, the Logo YOLO detection method achieved the mean average precision (mAP) values in the three major categories of clothing, food, and essential goods, with 61.58%, 56.42%, and 61.97%, respectively. The classification and detection performance of these two methods have significant advantages, providing effective technical support for the intelligent design of logos.

https://doi.org/10.5573/IEIESPC.2025.14.4.443

(Kai Zhao)

The research aims to eliminate the defects of image style transfer algorithm in image quality, content consistency and style consistency, and improve the generalization ability. An improved GAN image style migration algorithm AMS-Cycle-GAN is designed and implemented. In this algorithm, the generator uses the position normalization and moment shortcut modules, and the discriminator is based on the GAN model with channel attention mechanism and spectral normalization. The results showed that AMS-Cycle-GAN showed good performance in improving the visual quality, content consistency and style consistency of the generated image through the experimental verification under various settings. Especially in the photo2vangogh and vangogh2photo datasets, the IS and FID values were significantly better than those of other methods, reaching 5.3±0.8 and 114.1, and 4.94±0.45 and 148.23, respectively. The improved design of AMS-cycle-GAN generator and discriminator proves its value as a reliable and superior image style migration algorithm with its excellent performance.

https://doi.org/10.5573/IEIESPC.2025.14.4.457

(Qiufen Yu)

The ultimate aim of sports community service is to satisfy the growing demand for sports. Competitive sports have developed rapidly in the sports cause, so how to make better use of it to achieve the “feeding” of sports public service has become the focus of research. In this study, color clustering and image local entropy are combined to detect the plane scene, and the preliminary detection results are obtained. Then the foreground pixel is completed by using the target significance information to realize multi-target detection. According to the geometric constraints of the scene, a projection matrix solution method based on global optimization is proposed, and the sequential correlation strategy is applied to match the target points. The proposed model estimates camera motion parameters according to the functional relationship between feature points to realize 3D reconstruction of plane scene sequence. It is verified that the absolute and relative errors of the 3D reconstruction model are 2.11 mm and 0.42%, respectively. The average detection accuracy was 93.88%. It has good stability, applicability and reconstruction effect. The reasonable application of this model can promote the sports and enhance the enthusiasm of the whole people.

https://doi.org/10.5573/IEIESPC.2025.14.4.471

(Lu Tan)

This work proposes a novel approach for cross-modal graphic retrieval, leveraging deep learning and hash learning techniques. It aims to address the limitations of current multimodal information retrieval methods in capturing detailed information within individual modalities. Initially, a deep learning-based model is developed to extract features from text and image modalities. To further enhance the granularity of modality-specific information, a cross-modal hashing retrieval model incorporating graphic features is proposed. This model leverages attention mechanisms and adversarial networks to optimize performance. Experimental results demonstrate the effectiveness of the proposed model, achieving an average recall of 77.8% in graphic feature extraction with the highest classification precision of 0.637 and 0.712 on two separate datasets. Furthermore, the cross-modal hash retrieval model achieves an impressive average precision mean value of 0.833 in the image retrieval text task using a 64-bit hash code. These findings indicate that the proposed model surpasses comparable models in terms of precision-recall curve. The attentional mechanism, intermodal confrontation, and intra-modal confrontation modules significantly contribute to the model’s performance in image and text detection. Notably, the attentional mechanism module plays the most significant role, followed by the intermodal confrontation module. Consequently, this study’s model is well-suited for cross-modal graphic retrieval tasks.

https://doi.org/10.5573/IEIESPC.2025.14.4.483

(Yuan Ren)

To improve garden landscape spatial environment design effect, a sparse Bayesian algorithm based image visual reconstruction method is proposed to construct an optimized image visual reconstruction model. It is based on garden landscape spatial environment design. In the model construction, gray pixel feature separation method and multi-level feature decomposition method are combined for image visual feature reconstruction. The experimental results show that the minimum error value of the optimized sparse Bayesian algorithm is 0.080, the minimum running time is 264.5 s, and the maximum average PSNR and SSIM values are 25.941 dB and 0.715, respectively. The minimum resolution of the sparse Bayesian visual image reconstruction method is 408, the minimum error value is 0.0505, the maximum value and standard deviation are 0.123 and 0.0261, respectively. The average structural similarity of the sparse Bayesian image visual reconstruction method in 50 experiments is about 3.54. It improves by 54.6% and 88.3% respectively compared to convolutional neural based image reconstruction methods and non local variation based image reconstruction methods. The above results show that the image visual reconstruction method for landscape spatial environment design has certain application value in the field of landscape design.

https://doi.org/10.5573/IEIESPC.2025.14.4.495

(Jin Ma) ; (Wei Sun) ; (Yu Zhang)

To solve the traditional image classification issues focusing on local info while neglecting overall info and forgetting stroke info, and enhance natural image classification models, this study suggests a dual-channel, dual-path art image classification model. By setting up three primary colors and stroke information channels, combined with dual path networks and support vector machines, further extracting image feature information for classification. The experiment showcases that the dual channel constructed by the research institute has a minimum error rate of 42.13% for the Top-1 indicator in feature information extraction, which is 9.25% higher than the commonly used single channel Gram model in error rate accuracy. The error rate of the Top-1 index for art images of different styles, genres, and authors in the dual path network model is 11.73%, while the error rate of the Top-5 index is 0.73%. The dual channel dual path network model has a recognition and classification accuracy of 91.56% for different styles of art images, which is at least 1.25% higher than the commonly used image classification models. The experiment indicates indicate that the dual channel dual path network model constructed by the research institute has good classification performance.

https://doi.org/10.5573/IEIESPC.2025.14.4.507

(Junye Yang) ; (Yujuan Du) ; (Fang Liu)

In response to the problem of excessive reliance on the experience and knowledge of doctors in the current field of medical image processing, research has proposed two new algorithms: cell nucleus segmentation algorithms based on multi-task cascaded convolutional networks and medical image interpretable algorithms based on attention mechanism deep convolutional residual networks. The experimental results showed that the average F1 Score, IoU, DICE1, DICE2, Jaccard composite index, and average accuracy of the multi task cascaded convolutional network are around 0.84, 0.85, 0.88, 0.75, 0.81, and 0.80, respectively, which are higher than those of the DCAN and Jbarker models. The average diagnostic conclusion prediction accuracy, average diagnostic conclusion prediction recall, average semantic attribute prediction accuracy, and average semantic attribute prediction recall of the attention mechanism deep convolutional residual network were approximately 85.2%, 89.3%, 76.5%, and 79.9%, respectively, which are higher than the AlexNet and ResNet algorithms. In addition, the average diagnostic conclusion and semantic attribute prediction accuracy of the two indicators of the attention mechanism deep convolutional residual network were about 86.1% and 79.5%, respectively, which are higher than other models. The outcomes of the study indicate that both algorithms proposed can effectively process medical images.

https://doi.org/10.5573/IEIESPC.2025.14.4.520

(Prashant Upadhyay) ; (Tuhina Panda) ; (Preeti Jaidka) ; (Nidhi Gupta)

This paper presents AskAI, a unique approach to extractive product question answering from the product descriptions facilitated by a chatbot, designed specifically for E-commerce websites. AskAI employs a Chrome extension and a question answering model trained on a blend of SQuAD 2.0 (Stanford Question Answering Dataset v2.0) and ePQA (eCommerce Product Question Answering) datasets. The user inputs their query, and the system automatically scrapes relevant context from the webpage. Our model leverages transfer learning, utilizing a pretrained BERT model fine-tuned initially on the SQuAD 2.0 dataset and further trained on the ePQA dataset. This approach enables the model to effectively understand and respond to user questions. By integrating contextual information directly from the webpage, AskAI delivers natural-sounding and relevant answers. The results showcase the effectiveness of our approach, including an F1 score of 78.76 on SQuAD 2.0 and 72.31 on ePQA with an Exact match of 74.82 and 68.98 on SQuAD 2.0 and ePQA datasets respectively. Thus, demonstrating the capability of the model to accurately comprehend user inquiries and provide meaningful responses.

https://doi.org/10.5573/IEIESPC.2025.14.4.528

(Jin Young Choi) ; (Byung Cheol Song)

This paper proposes a novel synthetic dataset generation framework called Noise-to-Dataset to address data scarcity in semantic segmentation tasks on the LWIR domain. The framework consists of two stages: a denoising diffusion probabilistic model (DDPM) that generates semantic masks from Gaussian noise and a semantic diffusion model (SDM) that produces synthetic images based on these masks. Noise-to-Dataset enables the creation of diverse, high-quality synthetic datasets, significantly improving segmentation model performance. Experimental results show enhancements not only in LWIR datasets but also in RGB datasets like Cityscapes and ADE20K, highlighting its potential to generate valuable training data without the need for manual annotation.

https://doi.org/10.5573/IEIESPC.2025.14.4.535

(Xianfeng Zhu)

Transportation of agricultural products is an important part of agricultural development, but the problem of logistics center siting is a complex optimization problem. We adopt AI-based demand analysis and cold chain logistics center siting model based on improved ant colony algorithm, and improve the neighborhood structure and parameter settings of the algorithm, and carry out the process of initialization, iteration, and termination of the algorithm. We compare the algorithm in this paper with similar algorithms. The results show that the algorithm proposed in this paper is significantly better than the other five algorithms in both OBJ and ACC metrics, which indicates that the algorithm proposed in this paper has a stronger site selection effect and can better solve the site selection problem. In terms of practical application, the cold chain transportation time of agricultural products is reduced by shrinking, the total mileage of transportation is shortened, and the satisfaction level of each logistics center is significantly improved.

https://doi.org/10.5573/IEIESPC.2025.14.4.545

(Xiang Zou) ; (Yongfei Jia)

The digital economy, as one of the primary driving powers for contemporary economy, has had many negative impacts on the environment due to its vigorous development. The study aims to explore the key role and internal mechanism of digital economy development in promoting sustainable and harmonious development of China’s economy and natural environment. This study analyzes the impact of green transformation in the economic belt around the Yellow River using the distance method of superior and inferior solutions and neural networks. The results indicated that there was a disparity in the development of the digital economy in the Yellow River Economic Belt. This belt was mainly composed of digital environment, infrastructure, industrialization, and industrial digitization, which accounted for 42.34%, 28.87%, 16.87%, and 11.92%, respectively. There were significant distinctions in the digital economy among various provinces, with the highest being 0.6818 and the lowest being 0.1425. The general tendency was on the rise, but there was a decline in 2018. During the research period, the digital economy as a whole shifted towards the southwest, showing a trend of expansion in the north-south and contraction in the east-west. The study combines the distance method of superior and inferior solutions with neural networks to complete a win-win results of economic prosperity and green environmental protection in promoting the digital economy.

https://doi.org/10.5573/IEIESPC.2025.14.4.557

(K. Sarmila Har Beagam) ; (Mahabuba A) ; (Jayashree R) ; (Jennathu Beevi Sahul Hameed)

The standard single-slack bus power flow (PF) method has flaws in that the slack bus (mismatch) absorbs every loss. This means that the loss is being viewed in light of the slack bus. As a result, it is challenging to demonstrate the equity and transparency of the loss component. This study introduces a fresh way to clear and settle the single-sided auction market using the Honey Badger Algorithm (HBA) and a new method called Participant-Based Distributed Slack Reactive Optimum Reactive PF (PBDSOQF). In this model, the equality constraints are shown by the voltage levels at each bus, and the inequality constraints are shown by a single power balancing equation that matches the amount of power sent by all GENCOs and received by all DISCOs at the market center. It is shown that the suggested technique greatly minimizes the overall bus loss, lowering the objective function’s cost.