A GPU-free license plate detection based on fused color-edge and Retina approach.

With the great success of deep learning and IoT techniques, many methods with GPU for License Plate Detection (LPD) have attracted remarkable attention in recent times. However, GPUs are not always equipped for video surveillance equipment due to device costs in a distributed edge-computing environment. Thereby, most methods of LPD on edge devices have poor running performance without GPUs. In this paper, we propose a novel hybrid methodology for license plate detection, which can accurately and quickly locate license plates in complex visual surveillance scenes without GPUs. As the background colors in China's license plates are specific, blue or green, a well-designed color filtering and Sobel detector are used to preprocess images to eliminate the complex background, which rapidly speed up the process of license plate detection. Then, we employ heuristic Simple Moving Average (SMA) search method to filter out the license plate area that meets the definite conditions from the candidate areas. In addition, Support Vector Machine(SVM) and trained Retina model are introduced to further detect the position of license plate for pictures without any candidate areas to improve the accuracy. Finally, we evaluate the proposed method on Chinese City Parking Dataset(CCPD), which is China's largest public license plate data set. Results demonstrate that our method is 312.37% faster than the original Retina model on devices without GPUs and still achieves the precision of 97.95%. These results show that the proposed method outperforms other methods on detecting Chinese blue license plates on CPU devices. [ABSTRACT FROM AUTHOR]