Estimate the Minimum Quality of the Image to Recognize a Sri Lankan Vehicle Number from a CCTV Footage

Abstract

Identifying vehicles by the number plate is widely using in forensic applications to investigate criminals. The most prominent approach is using computer vision embedded software and highresolution cameras to detect and recognize vehicles in real-time. Due to the higher price of such systems, the widely using surveillance camera systems consist of low-resolution cameras. Highquality CCTV footage allows converting the blurred image into a clearly defined image in the society, which is beneficial for the development of the community. Trade experienced can be improved through using the CCTV conversion under the society development. The trade-off is the lack of pixel information stores for each image. Key issues related to image quality include blur, image aspects, camera positioning, light reflection, resolution, and night vision aspects. It has been negatively affected by forensic surveillance image analysis. The majority of the surveillance video footage received by the Government Analyst’s Department in Sri Lanka is degraded considerably and quickly unusable. The major drawback of analyzing images to recognize vehicle number plates is that there is no proper quality standard for the video footage. Hence time, resources, and human resources wasting over unusable images is a significant issue. This research study introduces a systematic way to recognize vehicle numbers in surveillance camera images without wasting time and effort. The main objective of this research analysis is to predict the recognizability of a vehicle number plate, which can be recognized by the custom trained Convolution Neural Network. Faster Region-Based Convolutional Neural Network (Faster-RCNN) trained with more than a hundred characters collected from surveillance camera footage of Sri Lankan vehicles have been used as the number recognition model. Based on the plate resolution, plate sharpness, and the recognized result received from the number recognition model, predictive analysis has conducted to determine the levels of resolution and sharpness that would support recognizing at least a single character of the number plate. The Logistic Regression Model was able to predict images with 70% of accuracy. The predictive analysis showed that as the resolution increases, the recognizability also increases. The image resolution is the most important quality attribute than the sharpness to identify characters using the number recognition model. The analysis also showed that increasing the sharpness has no significant effect on increasing the recognizability. Hence, it is more important to focus on resolution enhancement techniques over sharpness enhancement when preprocessing the number plate.