Research On Single License Plate Image Super-resolution Reconstruction Algorithm

In the field of computer vision,images are the carriers of many important information.Image quality determines the difficulty and effect of various visual tasks directly.Limited by the performance of imaging equipment and remote shooting,the quality of the image becomes poor,and the license plate characters cannot be directly recognized.Improving the quality of imaging equipment costs a lot,and there are still remote scenes in the image,so the super-resolution reconstruction of images by software has been widely applied.Since convolutional neural network can learn the complex mapping relationship between low-resolution images and high-resolution images,deep learning method to super resolve high-quality license plate images from the low-resolution license plate images is a hot topic in computer vision.The main task of this paper is to reconstruct the low-resolution license plate image into the corresponding 4× high-resolution license plate image.The main contents and innovations of this paper are as follows:1.In order to make full use of the high frequency information from the image and reduce the reconstruction error of the license plate image,this paper proposes a dense-connected residual sampling network(DRSN),which combines the residual structure and dense-connected structure.In the DRSN,a residual sampling block with sampling layers and the channel attention mechanism is constructed,which can better explore the high-frequency information of high-resolution features after sampling by using the channel attention mechanism.In this way,it is more conducive to the reconstruction of details such as the shape edge of the license plate image.In the global structure of the network,residual sampling blocks are dense-connected to improve the feature utilization;the 1×1 convolution layer is used to reduce the feature dimensions before every residual sampling block to purify the information and reduce the feature redundancy.Experiments show that the proposed algorithm can reduce the number of network parameters and running time,improving the performance of super-resolution reconstruction algorithm.2.In order to reduce the amount of data processing of the super-resolution network and improve the quality of the license plate,we firstly detects and extracts the low-resolution license plate area in the image,which provides a new idea for the super-resolution reconstruction of the interested area in the video surveillance.Then,the 4× sampling layer in the DBPN(Deep back-projection Network)algorithm requires a larger convolution kernel,resulting in slow convergence and sub-optimal results,this paper decompose the 4× sampling layer in the DBPN into two 2× sampling layers,so as to complete the iterative back-projection in a gradual way.The 2× sampling layer uses a smaller convolution kernel,which speeds up the convergence speed and improves the super resolution performance.The middle size features can provide more information for each iteration.In the gradual back-projection network,the features with the same size are fused by skip connections to improve the feature utilization;the 1×1 convolution layer is used to reduce the dimension of fused features,from which it reduces complexity of the network and preserves the key information.Experiments show that the algorithm in this paper can effectively reduce the data processing of super-resolution network and improve the clarity of license plate image and recognition of characters.