Enhanced Automatic Number Plate Recognition Using Yolov7, Resnet, and Image Rectification Techniques

ANPR (Automatic Number Plate Recognition) has revolutionised crime detection and traffic management. These days, systems that still rely on older versions of the YOLO (You Only Look Once) algorithm are outdated, slow, and sometimes replaced by newer computer vision technology, or a “basic machine learning approach” that is more accurate and reliable. This research enhances YOLOv7 by segmenting images into regions using an N x N grid. This research streamlines the process of partitioning images into bounding boxes and detections, specifically targeting tighter border cases. This research also exploits ResNet (Residual Network Architecture), which allows residual functions to learn easy identity mappings using layer inputs, resulting in improved learning capacity. Furthermore, the method incorporates an Image Rectification algorithm to rectify images affected by snow, dust, and low light (during bad weather). In this study, propose a convolutional neural network model specifically for Indian number plate detection, which train and validate on the MNIST digit dataset and large real-world Indian number plate datasets captured in various lighting conditions, both daytime and nighttime, using a camera. The system maintains a high level of reliability and processing efficiency by leveraging advanced hardware (NVIDIA RTX 3060) that takes approximately 25 seconds to process an image and decode its number plate. The experimental analysis outcomes illustrate model advantages over existing methods.