DEVELOPMENT AND DATA ANALYSIS OF A DEEP LEARNING-POWERED IoT SYSTEM FOR REAL TIME SMART NUMBER PLATE DETECTION IN FOGGY WEATHER CONDITIONS

Abstract

Number plate recognition of automobiles is a crucial aspect in smart transport systems supporting automatic vehicle tracking and identification. License plate recognition during foggy and hazy weather conditions, however, still poses a challenge due to visibility issues and fuzzy image features. The objective of this research is to suggest an in-real-time number plate recognition system by deep learning for identification under adverse weather conditions with datasets collected using IoT devices. The proposed approach is three-pronged: image dehazing via dark channel prior algorithm for enhancement as step one, license plate detection through a CNN-based object detection model, and recognition of characters using Optical Character Recognition (OCR) techniques with super-resolution enhancement. Although in this instance the IoT system is not being created, for simulation of real-world conditions, the research utilized IoT-source image databases. Experimental results show that high identification rates were achieved even in heavy haze and fog by the proposed approach, which attests to its effectiveness under adverse weather conditions and future implementation in smart surveillance systems. The technique is a low-cost, flexible method that can be trained and tuned for many real-time traffic and security tasks, particularly in areas most affected by poor weather. In addition, the method reduces reliance on human observation, maximizes operational effectiveness, and can be adapted to existing traffic infrastructure with minimal change. Its flexibility also makes it possible to be coupled with cloud-based solutions for continuous learning and remote monitoring to ensure maximum long-term system performance and responsiveness. Its modularity ensures that upgrading and enhancing is effortless, and the compatibility flexibility in a broad spectrum of image sources makes it highly versatile. Overall, this project promotes the creation of smart traffic management through a reliable license plate recognition solution for harsh environmental-conditions. The accuracy of the model at 20 Epochs is 93% and with trained on 50 epochs it reached to 96%.