Saputra, Vriza Wahyu (2023) Pengaruh Augmentasi Cuaca dan Pra Proses Citra pada Deteksi dan Pengenalan Pelat Nomor menggunakan YOLOv7 dan Pengenalan Berbasis Sekuensial. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6025211022-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6025211022-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2025. Download (25MB) | Request a copy |
Abstract
Deteksi dan pengenalan pelat nomor telah secara masif diterapkan di berbagai negara dalam penegakan hukum dan tata tertib lalu lintas. Namun, dalam praktiknya terdapat beberapa tantangan dalam proses mendeteksi dan mengenali pelat nomor tersebut. Misalnya, kamera CCTV yang beresolusi rendah, pencahayaan yang buruk dan kondisi lingkungan yang berbeda seperti siang, malam, hujan, panas dan berkabut yang membuat hasil deteksi pelat nomor tidak akurat. Pengambilan citra pada kondisi hujan dan berkabut sangat sulit dilakukan karena kondisi tersebut menyesuaikan dengan keadaan alam. Augmentasi tradisional seperti brightness increasing, color increasing, auto contrast, rotate, shear dapat mememperkaya dataset. Selain itu, Augmentasi cuaca dapat memperkaya dataset hujan dan berkabut sehingga model yang dilatih akan lebih handal. Pada dataset dengan kondisi pencahayaan yang buruk, diperlukan teknik untuk meningkatkan citra sehingga model yang dilatih akan lebih handal. URetinex-Net meningkatkan citra dengan kondisi pencahayaan yang buruk dengan memecah citra menjadi lapisan reflektansi dan pencahayaan sehingga citra menjadi lebih cerah. Pada deteksi pelat nomor, YOLOv7 adalah metode deteksi single stage yang akurat berfokus pada kecepatan serta akurasi. YOLOv7 menghasilkan MAP 56.8% lebih unggul daripada metode deteksi lainnya. Pada pengenalan pelat nomor, penulis mengadopsi beberapa metode Scene Text Recognition (STR) seperti CRNN, TRBA, ViTSTR, ABINet dan PARSeq. Hasilnya, deteksi pelat nomor menggunakan YOLOv7 mengasilkan MAP tertinggi sebesar 96.3%. Pada pengenalan pelat nomor, proses augmentasi + cuaca meningkatkan akurasi pada metode TRBA sebesar 3.68% daripada metode augmentasi tradisional. Akurasi tertinggi terletak pada metode PARSeq dengan hasil 90.56%. Proses peningkatan citra menggunakan URetinex-Net cukup berpengaruh terhadap hasil akurasi dengan akurasi terbaik pada metode PARSeq yaitu 90.42%.
======================================================================================================================================
Automatic license plate detection and recognition have been extensively implemented in various countries for law enforcement and traffic regulations. However, there are several challenges in detecting and recognizing these license plates. For instance, low-resolution CCTV cameras, low lighting conditions, and different environmental conditions such as day, night, rain, and fog can result in incorrect license plate detection. Capturing images in rainy and foggy conditions is particularly difficult as the conditions adapt to the natural surroundings. Traditional augmentations such as brightness increasing, color increasing, auto contrast, rotation, and shear can enrich the dataset. Additionally, weather augmentation can enhance the rainy and foggy datasets, resulting in a more robustly trained model. In datasets with low lighting conditions, techniques are required to enhance the images, thereby improving the reliability of the trained model. URetinex-Net enhances images with poor lighting conditions by separating the image into reflectance and illumination layers, resulting in brighter images. YOLOv7 is an accurate single-stage detection method for license plate detection that emphasizes both speed and accuracy. YOLOv7 achieves a superior mean average precision (MAP) of 56.8% compared to other detection methods. We adopted several Scene Text Recognition (STR) methods for license plate recognition, such as CRNN, TRBA, ViTSTR, ABINet, and PARSeq. The results show that license plate detection using YOLOv7 achieves the highest MAP of 96.3%. In license plate recognition, the augmentation + weather process improves the accuracy of the TRBA method by 3.68% compared to traditional augmentation methods. The highest accuracy is obtained with the PARSeq method, yielding a result of 90.56%. The image enhancement process using URetinex-Net significantly affects the accuracy, with the best accuracy obtained with the PARSeq method at 90.42%.
| Item Type: | Thesis (Masters) |
|---|---|
| Uncontrolled Keywords: | Scene Text Recoginition (STR), Deteksi Pelat Nomor, Pengenalan Pelat Nomor, YOLOv7, Augmentasi Cuaca, URetinex-Net, License Plate Recognition, Scene Text Recoginition (STR), YOLOv7, Weather Augmentation |
| Subjects: | Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis |
| Depositing User: | Vriza Wahyu Saputra |
| Date Deposited: | 23 Jul 2023 15:07 |
| Last Modified: | 23 Jul 2023 15:07 |
| URI: | http://repository.its.ac.id/id/eprint/98952 |
Actions (login required)
 |
View Item |