Automatic End-to-End Detection of Number Plate Using Neural Network for Indian Datasets

Developing Automatic Number Plates Recognition (ANPR) solutions for Indian vehicle number plates poses a significant challenge due to the diverse variations in size, font, script, and shape. A crucial requirement to address this challenge is a comprehensive dataset that encompasses the unique characteristics of the Indian scenario. However, the availability of such a dataset is currently limited, impeding the advancement made in developing ANPR solutions that are accessible to the public and can be easily reproduced. Unlike China and the United States, which have invested in creating comprehensive ANPR datasets such as the Chinese Citys Parking Dataset (CCPD) [1] and the Application oriented License Plates (AOLP) dataset [2], the development of an equivalent dataset specific to India has been lacking. In this study, we present a comprehensive dataset comprising 1.5k images [3], specifically curated to encompass a wide range of Indian number plates. This dataset serves as an invaluable resource for research in Automatic Number Plates Recognition (ANPR), particularly in the context of Indian conditions. To enhance the usability and applicability of this dataset, we propose a scalable and reproducible methodology for augmenting its contents. Using this expanded dataset, we delve into the exploration of an End-to-End (E2E) ANPR architecture that is specifically tailored for the Indian scenario. While the E2E architecture was initially introduced for Chinese vehicle number plate recognition using the CCPD dataset, our adaptation of this architecture to our Indian dataset has yielded significant insights, which are comprehensively discussed in this paper. This adaptation enables us to address the unique challenges posed by Indian number plates and paves the way for improved ANPR solutions in the Indian context. Our analysis highlights the challenges encountered when directly applying the E2E model from the CCPD authors to the Indian dataset. The substantial multiplicity in Indian number plates, along with differences in their distribution compared to the (ccp dataset), necessitates careful alignment between the innate of the Indian and Chinese dataset. Through this alignment process, we observed a significant improvement of 42.86% in license plate detection performance. Furthermore, we conduct a performance comparison between our E2E model and the YOLOv5 pre-trained model on the COCO dataset, Surprisingly, through the fine-tuning process using a collection of Indian vehicles images, we discovered that developing an ANPR solution for Indian conditions based on the COCO datasets proves to be more efficient than relying solely on the CCPD datasets, even when employing an equal number of Indian vehicle images for fine-tuning the detection module. Overall, this paper presents a valuable dataset tailored for ANPR research in the Indian context. Additionally, we provide insights gained through the customization of an E2E ANPR architecture. These findings underscore the necessity of tailored datasets and the importance of selecting appropriate pre-training datasets to achieve accurate and efficient ANPR solutions in the Indian scenario.