One of the most important parts of artificial intelligence projects made today is edge detection. Although it is used a lot in artificial intelligence studies, it is actually a difficult process due to reasons such as bad focusing and narrow dynamic range images, etc. In order to get rid of this kind of image problem, we created a unique filter for each image. In this paper, we provide a filtering method from eigenvalues and eigenvectors of the image matrix. Fundus images were obtained publicly from Drive, Stare, Messidor databases to test the performance of this method. After a lot of visual comparisons, we tested our filter and other filters using the Convolutional Neural Networks with confusion matrix to find out which filter better classify. 2, 5, 10 epochs were used for this process and the results were compared with 5000, 10000 and 15000 steps per epoch. Our filter was the best result with %69.23 accuracy with 2 epoch, 15000 spe. This result was followed by %65.38 accuracy Roberts filter with 10 epoch, 15000 spe and our filter with 10 epoch, 5000 spe.