Study On Super-Resolution Of SLP Image Based On Feature Fusion And Text Information

SLP(Ship License Plate),as the unique identification of a ship in the course of running,is of great significance for ship supervision,traffic record,evidence retention and SLP recognition.However,in the actual captured SLP image,due to the LR(Low Resolution),the quality of the captured SLP image is often poor,which increases the difficulty of ship supervision and recognition.Image SR(Super-Resolution)as the first choice solution to this problem has made great progress.However,there is still a lack of algorithms specifically aimed at improving the image resolution of SLP.The main problems include: 1)Lack of a reasonable process of generating SLP dataset? 2)The current SR network has insufficient ability to extract the features of SLP image? 3)Lack of a reasonable network structure in line with the SR requirements of the SLP image? 4)Lack of effective features extraction scheme for SLP text? 5)Lack of reasonable loss function for use of SLP text information.In order to solve the above problems,this thesis focuses on the image SR based on the whole and single row of SLP images,focusing on SLP feature fusion and text information extraction.Specific research work is as follows:1.PESRGAN(Parallel ESRGAN)algorithm based on parallel network and gradient supervision is proposed in this thesis to solve the problems of unreasonable dataset generation,inadequate image feature extraction and lack of text information.In this algorithm,an image degradation method is proposed to simulate the real scene,and the SLP dataset is generated.The parallel neural network is constructed to extract and fuse the deep feature and shallow feature respectively.Based on the loss function used in conventional SR networks,a gradient loss function is proposed which can be used to supervise the change of text edges,and can assist the network to update parameters by sharp text edges.Finally,the ablation experiment verifies the effectiveness of each improvement in the algorithm,and the comparison experiment verifies the advantages of the proposed algorithm over other algorithms.2.Aiming at the problems of current SR methods,such as poor extraction ability of text features,redundancy of SR network structure and ignoring font structure information,in this thesis,a SR algorithm TPSRGAN(Fusion Text Prior to Generative Adversarial Network for SR)based on text features and font structure is proposed.In the aspect of text feature extraction,text feature extraction is carried out by using the text recognition network pretrained in the SLP recognition dataset,and fused with the SR features,so as to enhance the ability of the network to extract the SLP text features.in the SR feature extraction module,the number of Residual in Residual Dense Block(RRDB)modules is streamlined to balance network performance and computing resource consumption.At the same time,the algorithm proposes the font structure loss function,which can guide the SR network to optimize the font structure in the single row SLP image.Finally,the ablation and comparison experiments are carried out on the SLP dataset,and it is verified that the proposed algorithm can effectively improve the image quality of single row SLP image.3.Developed a SR system for SLP image.With Qt Designer software as a development tool,the two algorithms proposed in this thesis are applied to the system,which can realize model selection,feature extraction,SR reconstruction,quantitative display of results and result saving,etc.The super resolution SLP image obtained by this system has higher resolution and quality.It can better assist the evidence retention,content identification and subsequent SLP recognition.