Enhancement Of Blurry Retinal Images Based On Deep Learning

Since capillaries,arteries and veins can be directly observed in the fundus retina,fundus images are widely used in the diagnosis,screening,treatment and evaluation of various cardiovascular and ophthalmic diseases,such as glaucoma,macular degeneration,diabetes retinopathy,hypertension,arteriosclerosis and so on.In recent years,telemedicine has been gradually popularized in order to reduce the gap between medical resources in remote areas and developed areas.As the main medium to assist doctors and computer diagnosis,fundus images play an important role in telemedicine.However,due to the requirements of hardware equipment and operation technology,the quality of fundus images obtained at the basic level is low,and most of them are underexposed.Properly reducing the light intensity of fundus photography can improve the patient’s medical experience,but it will also lead to problems such as insufficient exposure.Underexposed low light fundus images have some problems such as color distortion,blur,low contrast,high noise and so on,which need further image enhancement.As a preprocessing technology in the field of fundus image processing,fundus image enhancement can meet the needs of low light fundus image enhancement,and then achieve the double improvement of patient experience and remote medical accuracy.In the field of fundus image enhancement,the current mainstream research methods are traditional image processing methods,which can only enhance the image of a single fundus disease.The method based on deep learning can obtain stronger enhanced performance and versatility.However,there is a lack of research based on deep learning,and due to the limitation of data sets,most of the research focuses on cataract fundus image enhancement with paired image data sets,and there is a lack of research on unpaired fundus image enhancement.Therefore,Thesis proposes a self supervised deep learning network without paired images,which can enhance a single low illumination fundus image without paired fundus image data sets.The main innovations of this paper include the following two points.Firstly,thesis establishes a special data set for low illumination fundus images.As there is no low light fundus image data set at present,in order to facilitate the training of the deep learning network,based on the eyeq quality grading fundus image data set,this study screened the low light fundus images in the available tags,and made a data set specifically for low light fundus images for the first time,including 800 low light fundus images,which includes 600 training sets and 200 test sets.Each low light fundus image has a DR rating label labeled by a professional doctor.Secondly,thesis adopts a PE-GAN network based on attention mechanism and comparative learning mechanism.The U-Net with attention mechanism is used as the generator of GAN network.The global and local discriminators are used to distinguish the whole fundus image and local image blocks.At the same time,the contrast learning mechanism is introduced to simulate the normal fundus image to enhance the low illumination fundus image.Finally,the trained network can enhance a single low light fundus image.The obtained fundus image has good visual effects and practical diagnostic applications,and is superior to other enhancement methods in quantitative and qualitative analysis.It also has good enhancement effects on the low light fundus image caused by cataract.