2024, Vol.29, No.11

Journal of The Korea Society of Computer and Information 2023 KCI Impact Factor : 0.65

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

This study aims to establish a standard to accurately determine the similarity of the results when web pages are generated automatically using AI technology due to the explosive increase in demand for digital business. The YOLO, SSIM, Jaccard, and ORB techniques presented in previous studies related to the existing image similarity evaluation index generally focused on the partial and morphological similarity between the reference and the derived image. However, with the development of more complex and in-depth digital services based on generative AI, the need for comprehensive similarity analysis and determination methods that reflect the context and structure has emerged. Accordingly, this study proposed and verified a method to obtain ‘Web Structural Similarity (WSS)’ by combining the advantages of SSIM and ORB prior techniques. The research will serve various meaningful implications.

The rapid advancement of society due to industrialization, particularly through mass production enabled by automation, has led to the production of numerous products. However, it is difficult to ensure that all products are manufactured perfectly without defects. Therefore, identifying defects in products during the production process has become crucial. In modern society, detecting defects in various materials is highly valued. This paper focuses on detecting defects in plastic materials, which are among the most widely used and practical materials. In this study, we manually labeled the dataset, creating a dataset consisting of two classes. We utilized the YOLOv8 (You Only Look Once) model, which is capable of object detection, for training. To ensure fair evaluation, k-Fold Cross Validation was performed, resulting in an average F1 Score of 0.95, mAP50 of 0.97, and mAP50-95 of 0.68.

High-dimensional data causes difficulties in machine learning due to high time consumption and large memory requirements. In particular, in a multi-label environment, higher complexity is required as much as the number of labels. This paper proposes a feature selection method to improve classification performance in multi-label settings. The method considers three types of relationships: between features, between features and labels, and between labels themselves. To achieve this, a regression-based objective function is designed. This objective function calculates the linear relationships between features and labels and uses mutual information to compute relationships between features and between labels. By minimizing this objective function, the optimal weights for feature selection are found. To optimize the objective function, a gradient descent method is applied to develop a fast-converging algorithm. The experimental results on six multi-label datasets show that the proposed method outperforms existing multi-label feature selection techniques. The classification performance of the proposed method, averaged over six datasets, showed a Hamming loss of 0.1285, a ranking loss of 0.1811, and a multi-label accuracy of 0.6416. Compared to the AMI(Approximating Mutual Information) algorithm, the performance was better by 0.0148, 0.0435, and 0.0852, respectively.