應用深度學習演算法之海報文字區域檢測實驗

Abstract

近年來，數位化的廣泛應用也促使了互聯網的發展。伴隨著互聯網技術日新月異，大量的社交媒體和其他應用程式不斷推陳出新，數位圖像已然成為社會中一種主要的資訊獲取來源。在當今資訊量爆炸的社會裡，海報作為生活中最常見的資訊傳達媒介，成為生活中處處可見的藝術表現方式並充斥在現代人的生活當中。若能提出一個檢測方法來辨識海報中的文字區域，不僅能提取海報文字區域作為後續分析的資訊，也能使海報在網路中的更容易被使用者檢索。隨著深度學習的興起，越來越多研究者利用深度學習來完成影像分析及物件檢測。而其中，Mask R-CNN 與 Yolov4 分別代表著 two-stage 與 one-stage 的目標檢測方法，無論是在物件的瑕疵檢測、人臉的偵測、交通路況的偵測等領域都有很好的研究結果。然而，以上大多都是檢測自然場景物件，較少應用在平面設計的領域之中。基此，為了提取海報圖像的文字區域，本研究將訓練 Mask R-CNN 與Yolov4 兩個檢測方法，分別來對海報圖像文本進行檢測。實驗結果顯示，Mask R-CNN檢測文字區域的 mAP50 可達 79.0%；Yolov4 檢測文字區域的 mAP50 也高達 85.1%。意味著兩個目標檢測方法都可在海報版面中，定位出海報中文字區域，提供未來作為文字辨識的數據。而對比 Mask R-CNN 與 Yolov4 兩種演算法的輸出結果後，發現 Yolov4 可以更準確地檢測文字區域，並且較不受海報因色彩、文字大小、文字間隔等設計因素影響到檢測結果。

In poster design, designers often simplify and artisticize the information, quickly capture the audience's attention. The text in the poster must be brief and clear to the audience at a glance. If a detection method can be proposed to identify the text area in the poster, it can not only extract the text area of the poster as information for subsequent analysis, but also make the poster on the Internet easier to be retrieved by users. With the progress of deep learning and the improvement of computer hardware equipment, many researchers also use deep learning to complete image analysis and objectdetection. Among many object detection methods, Mask R-CNN and Yolov4 represent the two-stage and one-stage object detection methods respectively. Both of them have relatively outstanding performance in accuracy and computational efficiency. It can also be observed in real life that many researchers use this method to solve many problems, such as object defects detection, face detection, and traffic condition detection. However, most of the methods above detect objects in natural scenes, and are less used in the field of graphic design. In order to understand the ability of deep learning in poster layout analysis. In this study, two detection methods, Mask R-CNN and Yolov4, will be trained to detect poster image text respectively. The experimental results show that the mAP50 of Mask R-CNN can reach 79.0%; the mAP50 of Yolov4 can also be as high as 85.1%. It means that both object detection methods can be able to locate the text area in the poster layout, and provide data for text recognition in the future.