Earticle

본 연구는 2023년부터 2025년까지 전 세계 20개국의 iOS 빅데이터를 바탕으로 주요 생성형 AI 애플리케이션(ChatGPT, Gemini, Claude, Grok, DeepSeek)의 글로벌 시장 성과 동인과 현지화 전략을 실증적으로 규명하였다. 방법론적으로는 국가별 불변 특성을 통제한 패널 고정효과 모형 기반의 2단계 패널 분석과 Llama 3.1 및 Mistral-Nemo 모델을 결합한 고도화된 BERTopic 토픽 모델링을 수행하여 정량적 성과와 정성적 사용자 반응을 통합 분석하였다. 실증적 분석 결과, 사용자 평가의 양 적 지표(평가 횟수)는 다운로드와 매출액에 유의한 긍정(+)의 영향을 미치는 것으로 나타났으나, 기술적 고도화 단계에서 유입 증 가와 초기 불만이 공존하며 성과와 평점이 상충하는 ‘평점의 역설(Rating Paradox)’ 현상이 실증적으로 관측되었다. 특히 패널 분 석을 통해 글로벌 시장을 신호 민감형, 질적 가치 선별형, 파괴적/성숙 역설형, 시장 관성형의 4가지 유형으로 체계화하여 지역별 수용 패턴의 이질성을 입증하였다. 또한, 토픽 모델링을 통해 사용자 관심을 지능적 가치, 창의적 작업, 시스템 안정성, 운영 효율 성, 플랫폼 생태계, 메타 피드백의 6대 영역으로 분류하고, 정서적 유대와 인프라 안정성이 핵심 경험 요인임을 도출하였다. 본 연구는 고맥락 문화권의 정서적 현지화와 신흥 시장용 경량화 모델 전략을 제안함으로써, 생성형 AI 기업의 ‘글로컬라이제이션’ 확보를 위한 전략적 가이드라인을 제시한다.

This study empirically investigates the drivers of global market performance and localization strategies for leading generative AI applications (ChatGPT, Gemini, Claude, Grok, and DeepSeek) using iOS big data from 20 countries between 2023 and 2025. Methodologically, the research integrates quantitative performance with qualitative user responses through a two-stage panel analysis based on a fixed-effects model—controlling for country-specific invariant characteristics—and advanced BERTopic modeling integrated with Llama 3.1 and Mistral-Nemo models. The empirical results indicate that the volume of user evaluations (review counts) exerts a strong positive (+) influence on downloads and revenue. However, a "Rating Paradox" was observed, where market performance and average ratings conflict due to the coexistence of rapid user influx and initial dissatisfaction during stages of disruptive technological advancement. Notably, the panel analysis categorizes the global market into four distinct types—Signal-Sensitive, Qualitative Value-Selective, Disruptive/Mature Paradoxical, and Market Inertial—demonstrating the heterogeneity of regional adoption patterns. Furthermore, the topic modeling classified user interests into six domains: Intellectual Value, Creative Tasks, System Stability, Operational Efficiency, Platform Ecosystem, and Meta-feedback, identifying emotional bonding and infrastructure reliability as pivotal user experience factors. By proposing emotional localization for high-context cultures and Small Language Model strategies for emerging markets, this study provides strategic guidelines for generative AI firms to achieve successful "Glocalization."

This study examines the global implications of the United States’ withdrawal of support for digital trade norms, with a focus on developments since 2023. The research analyzes how this policy shift has affected the structure of international digital trade regulations and explores its impact on the evolving roles of key economies, particularly the United States and China. By comparing regional trade agreements (RTAs) and the dynamics of multilateral platforms such as the WTO Joint Statement Initiative (JSI), the study highlights the emerging fragmentation in digital trade governance. The absence of U.S.-led provisions in the final 2024 WTO e-commerce text reflects the challenges of multilateral digital rulemaking. The paper concludes by offering policy recommendations for reimagining international cooperation frameworks and identifying strategic pathways for third countries, including Korea, to enhance regulatory adaptability, digital competitiveness, and active participation in shaping future digital trade norms. These recommendations aim to support a fair, stable, and inclusive digital economic order amid shifting geopolitical and technological landscapes.

As the use of digital technologies such as IoT, AI, big data, and cloud computing spreads, the importance of data is increasing. Data is a key resource that determines the competitiveness of a company and the lifeline of the digital economy. Therefore, to activate digital trade, data must be freely transferred across borders. However, many countries are gradually strengthening regulations on cross-border data transfer to protect personal information. Therefore, to activate digital trade, it is necessary to find appropriate measures to protect personal information while ensuring cross-border data transfer. This paper examines major issues discussed in the field of digital trade, such as cross-border data transfer, data localization, and source code. We propose Korea’s response strategies to key issues related to cross-border data transfer, such as reorganizing Korea’s digital trade-related laws and regulations, improving digital trade-related policies and systems, actively responding to major countries’ digital trade policies, and utilizing digital trade agreements.

Digital trade refers to cross-border transactions based on data flows. These digital trade requires large amounts of data because trade transactions are conducted based on data, and cross-border data transfer plays a very important role in digital trade. Nevertheless, many countries are enforcing restrictions on cross-border data transfers for reasons such as privacy protection, cybersecurity, intellectual property infringement, and national security. These regulations on data transfer act as a significant obstacle to the development of digital trade. At present, not only does there not exist a unified international norm for cross-border data transfer, but even major countries have different regulations on it. Therefore, in order to facilitate digital trade, it is necessary to prepare measures to minimize regulations on cross-border data transfers and secure the interoperability of different data protection frameworks between major countries. Additionally, we will need to establish a global data privacy framework and conclude a global digital free trade agreement in the long term.

With the widespread use of powerful image editing tools, the demand for identifying the authenticity of an image is much increased. Copy-move forgery is one of the most common and immediate tampering attacks, and is one type of image forgery where one region of an image is copied to another region in an attempt to cover some potentially important features. In this paper a Novel approach is presented for image copy-move forgery detection and localization based on SVD and Projection Data. Experiment results demonstrate that our proposed algorithm can effectively detect multiple copy-move forgery and precisely locate the duplicated regions, even when an image was distorted by Gaussian blurring,JPEG compression and their mixed operations.

This study addresses the need to understand brand success by analyzing a crucial source of customer feedback: online reviews. We designed a comparative analysis to investigate the different factors that drive the success of two major fast-food brands, BBQ and KFC, in the competitive market. To achieve this, we combined several advanced analytical techniques, including text mining, sentiment analysis, TF-IDF, and CONCOR network analysis, to systematically process and evaluate their respective customer review data. Our results from the sentiment analysis showed that both brands garnered strong positive feedback, progressing from 'satisfaction' to 'recommendation' and 'best.' The TF-IDF analysis successfully identified key keywords for each brand, while the CONCOR network analysis revealed distinct competitive clusters: 'quality, localization, and service' for BBQ and 'price, convenience, and global' for KFC. These findings provide valuable insights into brand strategy. The effects of this study are the proposal of four major success factors— advanced localization, quality control, data capability strengthening, and local partnerships—which offer a strategic framework for businesses operating in this industry.

The purpose of this paper is to analyze the relationship between localization and performance of Korean corporations in Latin America. The empirical findings show that localization has a positive coefficient but statistical insignificance. Among the factors affecting corporate sales, only two variables such as introduction of new products and setting price have positive effects on corporate performances. And among variables affecting corporate net profits, only a variable of introduction of new products has a positive impact on corporate performances. JEL Classification: FO, F2, M1.

에드워드 스노든의 미국 국가안보국 활동 관련 폭로는 세계 여러 나라들이 자국의 이해관계를 지키기 위해 인터넷을 통제해야 할 필요가 있다고 생각하는 계기의 하나가 되었다. 물론 스노든 사태로 인해 모든 문제가 시작된 것은 아니고 하나의 개방된 자유로운 공간으로서 인터넷의 가능성과 실제로는 개별 국가 단위의 이해관계가 엇갈리게 마련인 현실에 기반한 존재로서의 속성 간에 내재한 긴장이 하나의 시대적 형상을 띠고 드러난 것이라고 이해하는 것이 합당할 것이다. 이 글에서는 이렇게 개별 국가가 여러 가지 이유에서 인터넷에 대한 통제를 강화하려는 시도를 이른바 인터넷 발칸화 현상이라는 관점에서 이해하면서, 국가 간 데이터 이전을 제약하려는 데이터 국지화 정책에 대해 인터넷 발칸화 논의의 맥락을 배경으로 그 위상과 의미를 검토한다. 이어 세계 주요 국가들의 데이터 국지화 정책의 동향에 대해 소개하고 이러한 정책적 근거의 당부에 대해 분석하며 나아가 이 문제의 법제적인 함의에 대해 어떤 시각과 자세로 접근하는 것이 좋을 것인가에 대해 논구한다.

Due to the disclosure of Edward Snowden on NSA, several countries including Germany and Brazil have been prompted to have attention on national security issue related to the Internet. USA has been maintaining the intitiative in the Internet industry for decades since its beginning. But the influence of USA on the Internet is not limited to economic area. Actually its prominent position exerts a decisive effect on a whole range of affairs - political, social, cultural, etc. - in other countries. Especially the national security issue has been the focus of interest in this context. So many came to think that the Internet had to be protected from foreign surveillance and invasion. The ideal of one, open, global network has been questioned. This phenomena can be described as the Internet Balkanization, and data localization policy can be considered one of its current symptoms. In this paper, global trend of data localization policy would be introduced, and then its legal implications would be examined. It is hoped that the appropriate measures to deal with these new changes could be found in legal and institutional dimension.

E-commerce would be defined as "commerce via the internet," in which commerce is the sale or purchase of goods or services. Short-run trade policies of e-commerce concentrating on B2C goods tend to be half-baked and need to be supplemented by policies towards services over the internet. Most of all, data localization issues will be at the heart of future trade policies regarding ecommerce. Korea is neither immune to, nor ready for data localization issues. It is time to construct a comprehensive panel dataset for e-commerce firms in manufacturing and services industries. Together with studies about the impact of e-commerce using such a dataset, Korea needs to make serious preparations to cope with digital localization issues with a balanced view.

This article builds a quantitative trade model to evaluate the welfare effects resulting from data localization measures. Based on the constructed model, it employs data from the World Input-Output Database to examine the welfare impact of data localization on the APEC economies. This paper highlights two distinct features of data localization, namely variable and fixed exporting costs. It shows that data localization measures can hurt all trading partners of the APEC economies in terms of real consumption. Further, the magnitude of the welfare effects stemming from data localization differs considerably across countries. This article suggests that the APEC economies should make efforts to minimize the welfare losses that arise from data localization measures.

Estimating the reliability of protein-protein interaction data sets obtained by high-throughput technologies such as yeast two-hybrid assays and mass spectrometry is of great importance. We develop a maximum likelihood estimation method that uses both protein localization and gene expression data to estimate the reliability of protein interaction data sets. By integrating protein localization data and gene expression data, we can obtain more accurate estimates of the reliability of various interaction data sets. We apply the method to protein physical interaction data sets and protein complex data sets. The reliability of the yeast two-hybrid interactions by Ito et al. (2001) is 27%, and that by Uetz et at.(2000) is 68%. The reliability of the protein complex data sets using tandem affinity purification-mass spec-trometry (TAP) by Gavin et at. (2002) is 45%, and that using high-throughput mass spectrometric protein complex identification (HMS-PCI) by Ho et al. (2002) is 20%. The method is general and can be applied to analyze any protein interaction data sets.

This paper presents a method to evaluate the performance of subpixel localization operators using target image data. Subpixel localization of edges is important to extract the precise shape of objects from images. In this study, each target image was designed to provide reference lines and edges to which the localization operators can be applied. We selected two types of moment-based operators: Gray-level Moment (GM) operator and Spatial Moment (SM) operator for comparison. The original edge localization operators with kernel size 5 are tested and their extended versions with kernel size 7 are also tested. Target images were collected with varying Camera-to-Object Distance (COD). From the target images, reference lines are estimated and edge profiles along the estimated reference lines are accumulated. Then, evaluation of the performance of edge localization operators was performed by comparing the locations calculated by each operator and by superimposing them on edge profiles. Also, enhancement of edge localization by increasing the kernel size was also quantified. The experimental result shows that the SM operator whose kernel size is 7 provides higher accuracy than other operators implemented in this study.

A computer vision technique to identify the location of an outdoor unmanned ground vehicle (UGV) is presented. The proposed technique is based on hybrid 3D registration of 360 degree laser range data to a digital surface model (DSM). Range frames obtained from 48 laser detectors are aligned with the reference coordinate system of the DSM. Three novel approaches are proposed for accurate and fast 3D registration of range data and the DSM. First, a two-step hybrid 3D registration technique is proposed. A pair-wise registration step of two consecutive range frames is followed by a refinement step using a layered DSM. Second, a fast projection-based pair-wise registration is proposed by employing rasterized 360 degree range frames. Third, a high elevation DSM is divided into several elevation layers and correspondence search is done near the vehicle's current elevation. This reduces the number of matching outliers and facilitates fast localization. Experimental results show that the proposed approaches yield better performance in 3D localization compared to conventional 3D registration techniques. Error analysis on five outdoor paths is presented with respect to ground truth.

The rapid development of shotgun proteomics is paving the way for extensive proteome profiling, while providing extensive information on various post translational modifications (PTMs) that occur to a proteome of interest. For example, the current phosphoproteomic methods can yield more than 10,000 phosphopeptides identified from a proteome sample. Despite these developments, it remains a challenging issue to pinpoint the true phosphorylation sites, especially when multiple sites are possible for phosphorylation in the peptides. We developed the Phospho-UMC filter, which is a simple method of localizing the site of phosphorylation using unique mass classes (UMCs) information to differentiate phosphopeptides with different phosphorylation sites and increase the confidence in phosphorylation site localization. The method was applied to large scale phosphopeptide profiling data and was demonstrated to be effective in the reducing ambiguity associated with the tandem mass spectrometric data analysis of phosphopeptides.

In this paper, we propose a probabilistic framework for model-based clustering of direction of arrival (DOA) data to obtain stable sound source localization (SSL) estimates. Model-based clustering has been shown capable of handling highly overlapped and noisy datasets, such as those involved in DOA detection. Although the Gaussian mixture model is commonly used for model-based clustering, we propose use of the von Mises mixture model as more befitting circular DOA data than a Gaussian distribution. The EM framework for the von Mises mixture model in a unit hyper sphere is degenerated for the 2D case and used as such in the proposed method. We also use a histogram of the dataset to initialize the number of clusters and the initial values of parameters, thereby saving calculation time and improving the efficiency. Experiments using simulated and real-world datasets demonstrate the performance of the proposed method.

본 논문은 지난 2016년 2월에 서명된 환태평양경제동반자(Trans-Pacific Partnership: TPP)협정을 분석함으로써 인터넷 제한조치 규제문제 관련 무역규범의 새로운 논의동향을 소개하고, 현재 협상 중에 있는 복수국간서비스협정(Trade in Services Agreement: TISA) 및 우리나라에 대한 시사점을 제공하는 것을 목적으로 한다. 최근 인터넷의 접근과 이용을 제한하는 기술과 법제도가 고도화되면서 글로벌 개방 인터넷의 분열과 지역화에 대한 우려가 증가하고 있다. 반면에 국가들의 인터넷 제한행위를 효과적으로 규제할 수 있는 구속력 있는 국제규범의 존재는 불분명한 상황이다. 이에 최근에 체결 또는협상중인 지역무역협정에서는 통상법적인 차원에서 과도한 인터넷 제한조치를 규제하기 위한 새로운 논의가 진행되고 있다. 특히 TPP협정의 전자상거래 챕터에 포함된 ‘전자적 수단을 통한 국경간정보이동’(제14.11조) 및 ‘컴퓨팅 설비의 위치’(제14.13조)에 관한 조항은 인터넷을 이용한 국경간의 정보이동을 제한하거나 데이터현지화 조치를 강제하는 인터넷 제한조치를 원천적으로 금지하고 있다. 또한 현존하는 WTO법을 적용함에있어서 발생할 수 있는 어려움과 우려사항을 우회하면서 인터넷 제한조치에 대한 통상법적인 대응을 효과적으로 전개할 수 있도록 고안되었다. 이러한 측면에서 TPP협정의인터넷 제한조치 관련규범은 TISA협상을 비롯한 지역적 차원의 무역규범을 비롯하여, WTO차원에서 진행되는 새로운 전자상거래 혹은 인터넷 무역규범 논의의 훌륭한 출발선으로 활용될 것으로 기대된다.

This paper purposes to introduce the new trend of the trade rules on internet restrictions by analyzing the Trans-Pacific Partnership (TPP) Agreement signed in February 2016, and to provide its implications for the Trade in Services Agreement (TISA) and Korea’s response to it. As the technologies and legal systems that restrict access and use of the internet have been advanced, concerns about the balkanization and localization of the global open internet are increasing. Nevertheless, the existence of binding international norms that can effectively regulate states’ internet restricting measures remains uncertain. Therefore, regional trade agreements, which have recently been concluded or being negotiated, are attracting attention as new discussions are under way to regulate excessive internet restrictions in terms of international trade law. Just to take the Electronic Commerce Chapter of the TPP Agreement as an example, in particular, it looks clear that Article 14.11 of “Cross-Border Transfer of Information by Electronic Means” and Article 14.13 of “Location of Computing Facilities” can contribute to prohibit internet restrictions which put up barriers to the free flow of information on the internet across borders. It is also conceivable that those provisions are designed to effectively deal with internet restrictions while circumventing the difficulties and concerns that may arise in applying existing WTO laws to the matter. In this respect, the new trade rules on the internet restrictions of the TPP Agreement are expected to serve not only as a good starting point for next regional or plurilateral trade agreements, including TISA, but also as a fine reference text to discuss new trade rules for electronic commerce and internet trade at the WTO level.

정부는 2019년 1월 금융 분야의 클라우드 이용 활성화를 위해 전자금융감독규정을 개정하였다. 하지만 클라우드 정책 및 규제들이 금융회사들의 자율적인 보안활동을 저해하거나 국민의 기본권을 일부 제한할 수 있어 금융권에서는 중요정보를 클라우드로 이용하려는 움직임이 거의 나타나지 않고 있다. 또한, 중요정보는 국내에만 두도록 하는 데이터 국지화 정책은 클라우드 이용 활성화를 가로막는 대표적인 규제이며, 이는 해외 사업자에 대한 차별문제도 야기시킨다. 따라서 데이터를 통해 디지털 금융혁신 기반을 제공할 클라우드를 활성화하기 위한 정책 및 규제 개선이 필요하다. 본 연구는 국내외 금융회사에 대한 클라우드 정책현황을 알아보고 국내 금융회사에 대한 정책 및 규제들을 분석하였다. 이를 통해 국내 금융회사에 대한 클라우드 정책의 한계와 문제점을 도출하고 금융회사의 클라우드 이용 활성화를 위한 데이터 국지화 규제 개선방안 등 정책적 대안을 제시하고자 한다.

In January 2019, the government revised the regulation on electronic financial supervision to revitalize the use of cloud in the financial sector. However, as cloud policies and regulations cloud undermine financial firms' autonomous security activities or restrict some of the people's basic rights, there has been little movement in the financial sector to use important information as the cloud. In addition, the data localization policy, which requires important information to be kept only in Korea, is a representative regulation that prevents the revitalization of cloud use, which also creates discrimination problems for overseas operators. Therefore, policy and regulatory improvements are needed to enable the cloud to provide a foundation for digital financial innovation through data. This study looked into the current status of cloud policies for domestic and foreign financial companies and analyzed policies and regulations for domestic financial companies. Through these efforts, the government aims to draw up limitations and problems in cloud policies for domestic financial companies and propose policy alternatives, such as measures to improve regulations on localizing data for financial companies to revitalize their use of cloud.

개인정보의 현지화(data localization)란 개인정보를 보관․처리하는 서버를 반드시 국내에 설치해야 하는 것을 말한다. 미 정보기관의 전세계적인 정보수집 활동이 드러남에 따라 많은 나라가 개인정보보호, 국가안보, 자국산업의 육성 등의 구실로 소스코드 공개를 요구하거나 데이터의 현지 보관 등 규제를 강화하는 경향을 보이고 있다. IoT, 빅데이터 시대에 정보의 자유로운 유통으로 경제 활성화를 기대할 수 있는 반면 많은 나라가 국가안보와 국익보호목적으로, 특히 EU의 경우 개인정보의 보호수준이 낮은 제3국으로의 정보이전을 불허하고 미국의 IT기업들이 플랫폼을 장악하는 것에 대한 경계심이 고조되고 있다. 일부 국가는 다국적 IT기업들이 엄청난 광고 수입을 올리면서도 세금을 거의 내지 않는 것을 막기 위해 국내에 데이터 센터를 두도록 유도하고 있다. 그러나 이러한 데이터 현지화는 관련업계의 클라우드 서비스의 이용이 늘어나는 추세에 역행할 뿐만 아니라 정보유통의 장벽이 되어 외국인투자 감소로 이어질 경우 GDP가 줄어들지 모른다. 주요국 현황을 보면 중국은 핵심 정보통신 기반시설 운영자에 대한 엄격한 책임과 의무를 부과한 네트워크 안전법에서 개인정보와 중요 데이터는 현지 서버에 저장하도록 의무화했다. 베트남도 OTT통달, 정령 72호에 의해 국내에 1개 이상의 호스트 서버를 설치하도록 요구하였다. 인도네시아는 공공서비스 전자시스템 관리자의 데이터 센터 국내 설치를 의무화하였고, 러시아는 개인정보는 국내에 있는 데이터베이스로 관리하고, 데이터 센터의 소재를 당국에 신고하도록 했다. 인터넷 시대에 데이터 현지화를 요구하는 법제는 나름대로 명분이 있어 보인다. 우리나라가 구글에 대한 공간정보 반출을 불허한 것도 북한이 장거리 미사일, 무인기를 날리고 있는 상황에서 부득이했던 것으로 판단된다. 그러나 공간정보는 유통․관광산업 및 자율주행차․드론의 이용에 없어서는 안 될 인프라이다. 인터넷의 발칸화 현상은 결코 바람직하지 않으며 국경간 정보유통에 대한 확고한 신념을 가져야 한다. 그 예외 사유는 안보 목적이든 무엇이든 엄격하게 해석하는 것이 타당하다.

Data localization requires personal data of citizens or residents be processed and stored within the border of the country. When such data are transferred overseas, they are required to meet local privacy or data protection laws. After revelations by Edward Snowden regarding United States global surveillance programs in 2013, an increasing number of countries have introduced data localization laws and data export restrictions. In this Age of IoT and big data, data localization laws worldwide are threatening digital globalization and inhibiting cloud computing’s adoption despite acknowledged benefits. There are various motivations for nations to adhere to data localization and restrict the transfer of personal data to a third country. Those restrictions are implemented for the purpose of national security, prevention of cybercrime, or protection of privacy of citizens, in particular, of the EU Member States. In some countries, it is to prevent the global IT Giants like Google and Facebook from evading local taxes on the huge advertisement profits in those countries by requiring the presence of local data centers. As such, local IT businesses are complaining of reverse discrimination from foreign providers of identical services. Other countries believe data localization offers a quick way to force high-tech economic activity to take place within their borders. It is also true that data localization and other barriers to data flows impose significant costs: reducing GDP by 0.7 to 1.7 percent in China, EU, Indonesia, South Korea, and Vietnam, which have all either proposed or enacted data localization policies. Statutory grounds in the above countries are found at: - China : Cybersecurity Law, the People's Bank of China Notice on Urging Financial Institutions to Protect Clients' Personal Financial Information (2011), and Measures for Administration of Population Health Information (2014) - Vietnam : Decree of Information Technology Services and OTT Circular -Indonesia : Organization of Electronic Systems and Transactions Regulation (82/2012) and Regulation No. 27 of 2015 regarding Technical Requirement of Equipment and/or Telecommunication Devices in LTE Technology Basis -European Union : Data Protection Directive (95/46/EC) and GDPR effective May 2018 -Russia : Federal Law 242-FZ -Korea : Act on the Establishment, Management, etc. of Spatial Data. There seem to be diverse rationales, one of which is national security or surveillance concerns. The Korean government is afraid of the detailed Google maps would be utilized by the belligerent North Koreans who have already developed long range missiles and unmanned aircraft systems. However, such mapping data are indispensable to local retail vendors, logistics and tourism enterprises and the operation of unmanned vehicles and drones in the near future. The so-called Internet balkanization and any barrier to free flows of data would hinder such burgeoning high-tech services. Therefore, the derogations of the primary aim of the Internet as an open network should be subject to strict scrutiny even for the compelling government interest such as national security.

최근 CAM[1]을 이용해서 이미지의 객체에 대한 주의 영역 또는 지역화(Localization) 영역을 찾는 방법이 WSOL의 연구로서 다양하게 수행되고 있다. CAM을 이용한 객체의 히트(Heat) 맵에서 주의 영역 추출은 객체의 특징이 가장 많이 모여 있는 영역만을 주로 집중해서 객체의 전체적인 영역을 찾지 못하는 단점이 있다. 여기서는 이를 개선하기 위해서 먼저 CAM과 Selective Search[6]를 함께 이용하여 CAM 히트맵의 주의 영역을 확장하고, 확장된 영역에 가우시안 스무딩을 적용하여 재학습 데이터를 만든 후, 이를 학습하여 객체의 주의 영역이 확장되는 방법을 제안한다. 제안 방법은 단 한 번의 재학습만이 필요하며, 학습 후 지역화를 수행할 때는 Selective Search를 실행하지 않기 때문에 처리 시간이 대폭 줄어든다. 실험에서 기존 CAM의 히트맵들과 비교했을 때 핵심 특징 영역으로부터 주의 영역이 확장되고, 확장된 주의 영역 바운딩 박스에 대한 Ground Truth와의 IOU 계산에서 기존 CAM보다 약 58%가 개선되었다.

Recently, a method of finding the attention area or localization area for an object of an image using CAM (Class Activation Map)[1] has been variously carried out as a study of WSOL (Weakly Supervised Object Localization). The attention area extraction from the object heat map using CAM has a disadvantage in that it cannot find the entire area of the object by focusing mainly on the part where the features are most concentrated in the object. To improve this, using CAM and Selective Search[6] together, we first expand the attention area in the heat map, and a Gaussian smoothing is applied to the extended area to generate retraining data. Finally we train the data to expand the attention area of the objects. The proposed method requires retraining only once, and the search time to find an localization area is greatly reduced since the selective search is not needed in this stage. Through the experiment, the attention area was expanded from the existing CAM heat maps, and in the calculation of IOU (Intersection of Union) with the ground truth for the bounding box of the expanded attention area, about 58% was improved compared to the existing CAM.