Automatic License Plate Recognition using Machine Learning for law enforcement applications

Among a variety of smart applications related to vehicle mobility, Automatic License Plate Recognition (ALPR) technology has achieved a diverse range of improvements over the recent decades. However, improvements are yet to be made to adapt to different purposes and circumstances. This paper introduces and discusses a novel and real-time ALPR framework that is adapted for law enforcement and surveillance purposes in developing countries, using Sri Lanka as the context. The study also aims to address weaknesses associated with existing ALPR systems. In this framework, a Convolutional Neural Network (CNN) is used to classify Regions of Interest (ROIs) extracted from vehicle images using contours as license plates (LPs) or non-license plates (non-LPs). A contour detection-based character segmentation algorithm is used to isolate the digits in LPs with an accuracy of 93.97%. In the final step, a Support Vector Machine (SVM) classifier with HOG features is introduced to identify the LP codes. A skew correction algorithm integrated with the Hough Line transform is proposed to perform the orientation corrections of tilted plates, and this algorithm proved to be efficient by significantly improving recognition accuracy up to 95.2%.