2.1. Computer Vision in the Media Industry

Processing photos or videos through various operations for specific purposes is referred to as digital image processing, which serves as a foundation for computer vision. Digital image processing digitizes images and represents them as functions, enabling various operations for machine learning and deep learning ^[1]. Based on these characteristics, computer vision technology dramatically influences the media industry, especially the overall value chain of video content production.

Computer vision and digital platforms are intricately linked in today’s innovative business environment. Incorporating computer vision technology into the media industry yields use cases across platform services, video content production, and functionalities embedded within smart devices. From a platform service perspective, computer vision technology has applications in image retrieval, personalized recommendation systems, and content generation facilitation. Leading portal platforms, such as Naver and Google, highlight this integration through services, such as Naver Smart Lens and Google Lens.

These offerings empower users to initiate searches based on the images they upload. This functionality extends beyond simple image search and is widely applied in e-commerce. Notably, OMNIOUS.AI stands out by delivering a proprietary AI-driven commerce solution encompassing OMNIOUS tagger, lens, and search components. This innovative solution caters to specialized fashion malls, such as KakaoStyle (ZIGZAG), ABLY, Hyundai Home Shopping, and Lotte ON. The technology effectively supports product searches and provides recommendations for similar products, significantly enhancing user experience and engagement. Additionally, computer vision technology extends to the webtoon industry, demonstrating its role in identifying and mitigating unauthorized distribution of creative work. Automating coloring processes through computer vision tools has significantly elevated the efficiency and productivity of webtoon creators, augmenting their artistic output and streamlining the creative workflow ^[7].

A representative example of computer vision technology applied in video content production is as follows. Disney employs factorized variational autoencoders for emotion analysis via facial recognition and to develop technology that predicts movie success based on this analysis ^[8]. Additionally, tvN, a leading content provider, used face editing technology–an AI application in computer vision–in the drama Navillera, producing results that were difficult for actors to replicate in person ^[9]. Recently, object-erasing technology has been used with image segmentation, and the generative adversarial network (GAN) has been introduced in film, drama, entertainment shows, and movies to erase logos and institutional names ^[10].

Finally, in smart devices, computer vision is integrated into core functions, offering diverse capabilities that have already become part of everyday life. For instance, functionalities include extracting textual content from images, applying correction filters and composition features during photo capture, and addressing problems, such as shaky hands. Notably, in the recent release of iOS 16, Apple introduced a novel feature that allows users to extract specific objects from videos and photos, enabling their insertion into messages, notes, emails, and more. Similarly, Samsung’s recent Galaxy series incorporated the photo remaster feature, which enhances low-quality photos with minimal degradation, and the AI eraser feature, allowing users to seamlessly remove unwanted objects from their images ^[11].

Beyond these examples, computer vision has diverse applications across the media industry, often synergizing with deep learning technologies to create novel forms of digital content. Consequently, as discussions and explorations of computer vision expand within the media industry, its influence extends across multiple sectors, including media and video content production. This demonstrates both the practical and academic influence of discourse on computer vision trends, highlighting the technology’s significance and future trajectory. Therefore, this study aims to provide insight into the media industry by comprehending the global research trends associated with computer vision.

2.2. Research Trends from Literature Analysis

Several studies have used research articles to examine trends in a specific field. In engineering, article titles, abstracts, and keywords have been used to identify research trends and suggest implications for future research ^{[4, 12, 13]}.

Kang et al. conducted a systematic literature analysis using academic papers to examine research trends in machine learning applications for production lines, focusing on the digitization of the entire production process ^[12]. Semeraro et al. reviewed the literature on machine learning techniques in the context of human-robot collaboration ^[13]. Hoekstra et al. analyzed literature on training text-event prediction algorithms with natural language processing in healthcare, using research data from the healthcare domain ^[4].

As noted, prior studies have sought to identify research trends and suggest future directions using machine learning techniques. However, differences in analysis methods were observed depending on the characteristics of each research area. For example, Zhang et al. identified research trends using keyword network analysis with virtual reality application-related data in the architecture and engineering fields ^[14]. Avsar and Mowla identified emerging technologies with high potential for future applications in smart agriculture through a research trend analysis of IoT-based agricultural systems ^[15].

The preceding research endeavors have investigated trends and anticipated future directions in fields directly influenced by AI. In traditional bibliometrics, methods involving keyword analysis of documents, such as papers and reports, have been used to comprehend trends within specific domains ^[16]. More recently, with the advancement of data science and data analysis software, semantic network analysis (SNA) using big data analysis methods has emerged to uncover research trends in particular disciplines ^[17]. In addition, SNA automatically extracts words from text-based documents and analyzes their inherent meanings based on co-occurrence relationships between words ^[18]. Given its capability to reveal latent semantic structures and contexts embedded in text, SNA is a highly advantageous methodology for analyzing tendencies and patterns across various domains ^[19].

The preceding studies employed SNA to comprehensively grasp research trends in big data. Analyzing keyword centrality, they successfully discerned vital research areas and themes within big data research. Based on the literature review of computer vision, this study aims to identify overall research trends, categorize research areas, and determine trending keywords in computer vision research by year using keyword network analysis.

3.1. Data Collection

This study collected articles published between 2017 and 2022 from the Computer Vision and Pattern Recognition (CVPR) Conference, the European Conference on Computer Vision (ECCV), and the International Conference on Computer Vision (ICCV), which are leading global conferences in computer vision. CVPR, ECCV, and ICCV are widely recognized as the foremost international conferences in engineering, particularly in computer vision. These conferences serve as premier platforms for researchers and practitioners in computer vision and related domains to present, discuss, and advance cutting-edge research findings, ranking among Google Scholar’s top publications.

We used the Web of Science database to collect conference articles, which provided most of the data. Web of Science was used for this analysis due to its trustworthiness and reliability, as researchers widely use it to determine titles, countries, and affiliated organizations. This study included articles from the Web of Science database, most of which contained information on titles, countries, and affiliated organizations. The research collected conference paper data not covered by the Web of Science from the “Accepted Paper List” provided by each conference’s website for each year. Specifically, 15,328 papers were gathered from the CVPR, ECCV, and ICCV conferences.

3.2. Data Analysis Methodology

4.1. Annual Publications

Fig. 1. Annual publications.

The trend in the number of papers presented at major international conferences in computer vision from 2017 to 2022 is illustrated in Fig. 1 Research in computer vision has consistently exhibited a rising trend every year since 2017 (Fig. 1).

4.2. Semantic Network Analysis

4.2.1 Title

After refining the collected paper titles using VOSviewer, 28,198 keywords were generated. The most frequently appearing keyword was video (258 times), followed by image (199), end (177), object detection (166), wild (141), and semantic segmentation (136). One keyword was used more than 200 times, whereas nine were used between 100 and 200 times, and 22 were used more than 50 times. The study fields and tasks commonly used in computer vision research and learning methods and models employed in computer vision AI training were prominently observed (Fig. 2). Table 1 summarizes keywords used more than 50 times with their frequencies.

Table 1. Keywords Frequency Analysis

Frequency	Keywords
200 times and above ($n = 1$)	video (258)
100-200 times ($n = 9$)	image (199), end (177), object detection (166), wild (141), semantic segmentation (136), GAN (116), transformer (113), person re-identification (113), human pose estimation (101)
50 times and above ($n = 22$)	CNN (99), deep neural network (84), single image (80), object (77), motion (76), point clouds (73), unsupervised domain adaptation (68), depth (65), action recognition (65), translation (61), shot (61), dataset (60), segmentation (58), reconstruction (55), neural network (55), fast (55), shape (54), network (54), model (54), efficient (52), representation (51), autonomous (51)

Fig. 2. Semantic network analysis results of the article titles.

4.3. Affiliated Organization Analysis

An analysis was conducted on the affiliations of papers presented at international computer vision conferences from 2017 to 2022 (Fig. 4). The analysis revealed that university-affiliated institutions accounted for the majority (51.29%) of the papers, indicating that research conducted solely by universities was most prevalent. Furthermore, collaborations between universities and corporations ranked second in frequency, followed by joint research efforts between universities and government-affiliated research institutions.

The year-by-year trends in collaborative computer vision research across affiliations revealed notable patterns (Tables 3 and 4). Overall, collaborations between universities and corporations displayed a substantial increase. Additionally, joint research between universities, corporations, and government-affiliated institutions generally grew faster than research conducted by each institution independently. The result suggests a trend toward increased collaborative research endeavors rather than independent research efforts by these institutions.

Fig. 3. Semantic network analysis results of the coauthor countries.

4.4. Annual Keyword Frequency Analysis

4.4.1 Research field

Based on the categorized keywords, research related to images, videos, 3D, and multimodal language steadily increased (Table 5 and Fig. 5). During the period, there were 2,273 articles related to images, 1,380 to videos, 1,350 to 3D, and 335 to multimodal language. The number of research articles with these keywords more than doubled from 2017 to 2022. In particular, research on images grew rapidly after 2018, while other computer vision fields increased more gradually. Research on 3D followed a similar upward trend but stagnated in 2020. However, with the increase in research on videos, research on 3D again increased after 2021. Moreover, research on multimodal language, considered a distinct field, has increased since 2021.

Fig. 4. Affiliated organization analysis results; university (U), corporation (C), government-affiliated research centers (G).

Table 5. Annual changes in research field.

Keyword	Years
	2017	2018	2019	2020	2021	2022	Total
image	195	265	394	415	446	558	2273
video	131	169	193	228	304	355	1380
3D	91	137	182	252	298	360	1350
language (multimodal)	29	24	32	52	67	131	335

Fig. 5. Annual changes in research fields.

4.4.3 Learning method

Significant changes were observed in learning methods, including representation learning, deep learning, shot learning, unsupervised learning, semi-supervised learning, supervised learning, and reinforcement learning. The analysis highlights dynamic shifts in learning methods over time. Deep learning dominated in the early stages, but its prevalence shifted over time. Since deep learning is a broad term, more specific learning methods are now emphasized. Research on representation and unsupervised learning methods increased significantly, whereas deep and reinforcement learning research declined. Specifically, supervised, semi-supervised, and unsupervised learning increased until 2018, after which growth slowed. Unsupervised learning research increased rapidly after semi-supervised learning gained attention in 2019. This suggests that growth in semi-supervised learning boosted research on unsupervised learning (Table 7 and Fig. 7).

Table 6. Annual changes in tasks.

Keyword	Years
	2017	2018	2019	2020	2021	2022	Total
classification	124	139	157	211	194	211	1046
detection	89	125	211	255	283	350	1313
segmentation	73	104	174	204	263	290	1108

Fig. 6. Annual changes in tasks.

Table 7. Annual changes in learning method.

Keyword	Years
	2017	2018	2019	2020	2021	2022	Total
representation learning	18	18	50	73	111	118	388
deep learning	61	78	77	63	53	39	371
shot learning	14	18	49	46	51	64	242
unsupervised learning	22	27	45	55	101	102	352
semi supervised learning	2	8	12	25	29	32	108
supervised learning	8	16	13	19	18	15	89
reinforcement learning	9	21	9	22	7	6	74

Fig. 7. Annual changes in learning methods.

4.4.4 Learning model

Analysis of learning models revealed that research on transformers has increased dramatically since 2020. Hence, the studies on the CNN and RNN demonstrated a significant decrease after the rise of the transformer. In contrast, research on the GAN and DNN was steadily conducted until 2021 (Table 8 and Fig. 8).

Table 8. Annual changes in learning model.

Keyword	Years
	2017	2018	2019	2020	2021	2022	Total
convolutional neural network (CNN)	65	63	87	69	29	33	346
deep neural network (DNN)	39	62	65	71	64	54	355
graph neural network (GNN)	2	2	3	10	15	11	43
recurrent neural network (RNN)	7	7	4	3	1	1	23
generative adversarial network (GAN)	21	61	89	91	86	98	446
transformer	4	6	5	23	135	346	519

Fig. 8. Annual changes in learning models.