結合Unreal Engine 4與生成式對抗網路之車牌影像合成系統

Abstract

車牌辨識已經是一門成熟的技術，廣泛被應用於停車場之車輛管理、道路收費系統、道路交通監測等領域。隨著深度學習的蓬勃發展，現已有許多能夠出色完成車牌辨識之網路，然而在訓練中不可或缺之車牌資料，於資料之獲取及準備階段相當耗費時間跟人力資源，甚至根據地區或狀況上的差異，需要之車牌資料會有字母、字型、角度、環境等不同的組合，若符合欲研究條件之車牌資料不足，也會有需要重新拍攝合適之車牌資料並重新進行標注的可能。因此本論文提出了一種車牌資料的合成方法，結合3D圖像軟體Unreal Engine 4以及CycleGAN，嘗試在不使用任何真實車牌影像輔助之條件下製作出可使用於車牌辨識訓練之合成車牌資料，以減少車牌辨識之相關研究用於資料收集及標記整理之時間與人力成本並強化車牌辨識效果。合成出來之車牌影像於視覺上與真實車牌相似，並且被證明能有效地提升作為辨識網路之YOLOv4之準確度。以3179張之真實車牌影像所訓練出來之YOLOv4為比較基準，我們所提出之合成方法所製作出的車牌影像能將原本97.00%之mAP提升至98.04%。With the development of deep learning in image recognition, automatic license plate recognition is widely used in urban parking lots, roadside toll systems, and road traffic monitoring. In the study of automatic license plate recognition, we are faced with many challenges such as complex road environment, blurred license plates, diverse shooting angles, and scarce license plates. In order to train a powerful model, a sufficient license plate data set is essential. However, it is very difficult to build a sufficient license plate data set. In addition to sufficient quantity, it is also necessary to collect all kinds of license plate texts and images taken in various environments. This is a very labor-intensive and time-consuming task. Therefore, in this thesis, we propose to generate synthetic license plate images using Unreal 4 engine and CycleGAN. In order to reduce the time and labor cost of license plate collection, we intend to generate synthetic license plate data that can be used in training license plate recognition model without using any real data. The synthetic images we generated not only look similar to the real license plates, but are effective for training Yolo's license plate character recognition model. Taking the model trained by YOLOv4 from 3,179 real license plate images as a comparison benchmark, by adding the synthetic data we generate, the mAP of YOLOv4 can be increased from 97.00% to 98.04%.