Super-resolution Restoration of Single Vehicle Image Based on ESPCN-VISR Model

Abstract In the process of image acquisition monitoring in vehicle license plate recognition system, the accuracy of license plate recognition will decrease which is caused by the complicated application scenarios, poor performance of the equipment and system, and the degeneration of surveillance images. In order to overcome the shortcoming, a super-resolution restoration method for single vehicle image based on efficient sub-pixel convolutional neural network (ESPCN) model is proposed (hereinafter referred to as ESPCN-VISR method). In the experiments, ESPCN-VISR method was proved to be superior by four indexes of quantitative assessment which are peak signal to noise ratio (PSNR), structural similarity (SSIM), vehicle license plate recognition accuracy (VLPRA), and calculation time of reconstruction (CTR). Compared with the methods based on sparse dictionary learning and deep convolutional neural network SRCNN, the ESPCN-VISR method can improve the performance of vehicle license plate recognition system with vehicle license plate sample image dataset LPI-1000.