Research And Implementation Of Small Target Detection Algorithm In Fundus Image

Diabetic Retinopathy(DR)is one of the major causes of blindness and Micro-aneurysm(MA)is the earliest detectable micro-abnormality.At present,the diagnosis of micro-aneurysms depends mainly on the screening of fundus retinal images by professional ophthalmologists,which not only requires high professional skills of ophthalmologists,but also the diagnosis process consumes a lot of time and energy,and medical resources are increasingly scarce with the rapid increase in the number of DR patients.In recent years,the combination of human visual attention theory has achieved well results in image processing,which provides a good idea to solve the problem of low efficiency of manual screening.Therefore,this thesis carried out a series of research and implementation around the automatic detection of small target in fundus image,including target detection algorithm based on multi-layer attention mechanism and target confidence discrimination method based on morphology.The main work is as follows:(1)Research and implementation of data processing method for fundus image quality balance.Aiming at the problem of uneven quality and insufficient quantity of fundus images,the equalization of fundus image quality is realized through color space and histogram image enhancement theory,combined with threshold segmentation algorithm and sliding window method,and the problems of unobvious small target features and insufficient data are solved.Experiments on a variety of target detection models verify that the processing method is effective and feasible.(2)Research and implementation of target detection algorithm based on multi-layer attention mechanism.For the problems of most target detection algorithms only use the top layer features for prediction,while small target features are not significant,a method of feature self-selection based on attention receptive field is constructed,by selecting the feature layers with obvious features of small target,and then fusing multi-layer features based on attention mechanism,paying more attention to the useful information for the task,then realized the detection of micro-aneurysms and get performance improvements.(3)Research and implementation of target confidence discrimination method based on morphology.According to the morphological analysis of the position relationship between micro-aneurysms and blood vessels,the U-net model is used to segment the blood vessels accurately in fundus image,then calculate the distance between the blood vessels and the targets which were initially identified.According to the distance relationship,the confidence of the preliminary detection results was scored and the false positive candidate targets were second selected by majority voting.This method can detect more microaneurysms while ensuring accuracy.