Research On GAN-based Color Document Image De-illumination Method

Paper documents are an important medium for us to convey information and express our thoughts.They contain a lot of information.But over time,people find that paper documents have many drawbacks,such as not easy to save,difficult to retrieve and so on.With the advent of electronic devices,the concept of document digitization has gradually come to people's attention.Save paper documents as pictures on your computer for easy storage and retrieval.In the past,we used printers,scanners and other devices to store paper data,but now with the popularity of micro-electronic devices such as mobile phones,people tend to use such devices directly to obtain information.Documents captured with a mobile camera are usually more susceptible to light than scanners,because the light source is often hidden by the camera or the user's hand,and even if not,the photos on the document are usually uneven.In order to improve the quality and readability of captured documents,this paper studies the lighting removal of document images.In this paper,a method of light equalization based on generation for antagonistic network is presented for color document images taken by mobile phones with uneven illumination.By building a generative antagonism network with ResUNet++ as the generator,the synthetic data and the real data taken are combined as training sets to train,and the neural network prediction is used to achieve the problem of color document picture deillumination.To demonstrate the superiority of this method,this method is compared with the more popular methods in recent years.The test results show that this method has better effect in the field of illumination and stronger generalization ability of the model.Lastly,there is a web-based document delighting system in my paper for user experience.The main work of this paper is as follows:(1)In order to solve the shortage of training data and the weak generalization ability of the model,a method of generating uneven illumination document based on CycleGAN is presented.By designing a generation transformation formula based on the instantaneous parameters to generate illumination background image and using CycleGAN to expand random information,the generated illumination background image is merged with the original image.This can be used to manually generate a large number of uneven lighting document pictures that simulate real lighting scenes,thereby improving the richness of the data.The test results show that the training model of the dataset generated by this method is more robust and deilluminates better than using the uneven image synthesized directly from the illumination background image as the dataset.(2)A color document delighting model using ResUNet++ as generator is presented.In view of the insufficient consideration of the dependence of the existing document image delighting model on the image channels in the learning process and the reduced performance of the model on color document images,a method using ResUNet++ as generator is proposed.ResUNet++ is composed of residual module,extrusion and excitation module,Arous space pyramid pool(ASPP)and attention module.The residual structure can be used to construct a deeper network model.The spatial pyramid pool is used to obtain multiscale information of the image,and the attention mechanism is introduced to highlight the image characteristics.The problem of model degradation can be solved while the correlation information between channels can be obtained,which can improve the interdependence between channels,thus improving the image processing performance of the model.(3)This paper has designed and implemented an online image processing system.I use both Vue and SpringBoot core technologies in my paper.The image de illumination service is implemented by pytorch framework of Python language.At the same time,in order to realize large-scale text image de illumination task,Kafka message queue is used as asynchronous communication component between back-end web service and de illumination service.The stability of the system is guaranteed.