Research On GAN-based Weakly Supervised Text Image Deblurring

With the rapid rise of computers and mobile devices,text images and photos carry more and more important information.However,the problem of quality degradation of text images is becoming more and more prominent,such as image quality degradation caused by image blurring,image smearing,camera shake,and image compression.This article mainly solves the problem of text image blurring in the degradation problem.In recent years,due to the rapid development of deep learning,in particular,with the rapid development of computer vision,there are more and more methods to solve common image degradation problems based on deep learning,but there is still a lack of a reliable method in the field of text image deblurring.Solution and related Chinese and English data set support.Based on Generative Adversarial Networks(GANs)and Transfer Learning,this paper proposes a weakly supervised generative adversarial recurrent network and a method based on Backbone Networks and generative adversarial learning deblured network.The main work content is as follows:(1)This paper proposes a weakly supervised text image deblurring method.Based on the confrontation generation network and the cyclic confrontation neural network,a reliable text image deblurring method is proposed,which realizes the migration learning from natural scenes to text image scenes through migration learning.In particular,through the loss function and model design method proposed in this article,you can Realize a stable weakly supervised model training process and achieved better results on the public data set.Compared with the original CYCLEGAN network,the loss function and specific network design in this paper have greater advantages.Similarly based on the previous method,the cyclic adversarial neural network implemented in this article is a process that can not only complete the deblurring operation,but also generate a blurred image through a clear image,and form a stable rising effect through the process of mutual supervision,and the generated image The blur effect is different from the synthesized picture and is similar to the deblurred text image trained in the real scene.More importantly,by using the data generation method proposed in this article,a larger text image deblurring can be achieved during the training process,and the actual data effect can be increased,that is,a large amount of simulation data can be synthesized from a small amount of real scene pictures.(2)This paper proposes a text image deblurring method based on adversarial neural network.Based on the adversarial neural network,the generator and loss function of the neural network are improved.It is proposed to use the replaceable Backbone and FPN network structure and perform well on the existing data set and self-labeled data set.It is in the public data set.Text CNN surpasses the existing public models,and the use of lightweight models can shorten the network's inference speed.(3)Efficient application deployment and visualized results display.This paper efficiently deploys and runs the model on the existing efficient network structure and the inference engine Tensor RT proposed by NVIDIA,which can visualize the results of the computer vision neural network model trained and processed in the laboratory in a short time,Intuitively show the excellent effect of model output.And through the replaceable way,the model can be replaced online,which greatly saves the time of the laboratory researcher to focus on the exploration of the model.