運用深度學習模型產生品牌廣告圖

Generating Brand Advertising Images by Deep Learning Models

李陽昇(Yang-Sheng Li)

指導教授 ： 李瑞庭

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102314

摘要

許多公司利用社群平台貼文打廣告以提升品牌形象與知名度，但貼文的生命週期短，為了提升使用者參與度，公司須頻繁地推出新的貼文，而大部分的貼文中都包含廣告圖，但這些廣告圖的製作主要是靠人力完成，既費時亦費力。因此，為企業開發一個能夠自動生成廣告圖的工具是不可獲缺的，而且工具最好能夠很直觀、很容易使用，如：透過文字描述廣告圖的內容就能產生對應的廣告圖。但就我們所知，目前未有研究專注於幫企業自動生成廣告圖。因此，在本研究中，我們提出了一個研究架構，幫助企業從文字描述自動生成廣告圖。我們提出的研究架構包含四個階段，第一階段，我們解析輸入的文字描述，並導出結構化表示；第二階段，我們提出了一個布局模型，利用第一階段導出的結構化表示，規劃每個物體的布局；第三階段，我們分別採用背景生成模組生成背景，物體生成模組生成廣告圖中的每個物體；最後，我們提出了一個渲染模組，將生成的布局、背景和物體渲染成廣告圖。實驗結果顯示，我們的方法優於比較方法，並能從文字描述生成真實且和文字描述語意一致的廣告圖。本研究可幫公司自動生成廣告圖，亦可幫廣告設計師產生設計圖，進而降低成本與節省時間。

關鍵字

廣告圖生成 ； 文字轉圖像生成模型 ； 布局模型 ； 深度學習 ； 網路廣告

並列摘要

With the increasing popularity of online social networks, many companies often make posts to promote brand image and awareness through their business accounts on social media platforms. However, the posts usually have a very short lifespan. To increase user engagements, companies might need to frequently make posts, which usually contain some advertising images. The design of advertising images is mainly done by manpower, which is time-consuming and costly. Thus, developing a tool for companies to automatically generate advertising images is both desirable and essential. It is better if the tool is easy and straightforward to use such as using a text description to describe the content of an advertising image. To the best of our knowledge, there is no study dedicated to automatically generating advertising images for companies. Therefore, we propose a framework to generate an advertising image for a company from a text description. Our proposed framework contains four phases. First, we parse the input text description to derive the structural representations. Second, we develop a layout model to plan the layout of each object by using the derived structural representations. Third, we employ the background generation module to produce the background, and the object generation module to produce each object to be presented in the advertising image. Finally, we develop a rendering module to render an advertising image by using the generated layout, background, and objects. The experimental results show that our proposed framework outperforms the compared methods, and can generate realistic and semantically consistent advertising images from text descriptions. The proposed framework can help companies generate advertising images and assist advertising designers in creating images in the design processes, which results in cost reduction and time saving.

並列關鍵字

advertising image generation ； text-to-image synthesis model ； layout model ； deep learning ； online advertising

參考文獻

Alalwan AA (2018) Investigating the impact of social media advertising features on customer purchase intention. International Journal of Information Management 42:65–77.

Braun M, Moe WW (2013) Online display advertising: Modeling the effects of multiple creatives and individual impression histories. Marketing Science 32(5):753–767.

Burt PJ, Adelson EH (1987) The Laplacian pyramid as a compact image code. Readings in Computer Vision. Morgan Kaufmann, San Francisco, CA, 671–679.

Dehghani M, Tumer M (2015) A research on effectiveness of Facebook advertising on enhancing purchase intention of consumers. Computers in Human Behavior 49:597–600.

Denton E, Chintala S, Szlam A, Fergus R (2015) Deep generative image models using a Laplacian pyramid of adversarial networks. Proceedings of the International Conference on Neural Information Processing Systems. 1486–1494.