چكيده به لاتين
The progress in the field of retinal imaging over recent years has significantly improved. This progress has given rise to presenting automated diagnosis systems for eye diseases. Among all retinal images, retinal fundus image is extensively used to diagnose eye diseases. A huge proportion of this image is covered by retinal blood vessels. These vessels disrupt the process of segmenting various lesions and also disease classifications based on the underlying structure. Therefore, it is necessary to remove these vessels in the preprocessing stage of the proposed diagnostic system so that the required information from important regions of the image like macula, optical disk, and etc. can be available for further processing. After specifying the location of blood vessels, they can be considered as unknown regions called as image mask that their values should be restored.
Restoration in the image processing terminology is carried out many methods referred to as inpainting. One of these methods is sparse-based inpainting. Typically, in Sparse-based image inpainting, a dictionary according to undamaged image patches is trained using dictionary learning methods. Having this in mind that sparse representation in the known and unknown regions is the same, the sparse coding problem is solved, and the specified regions will be inpainted by the usage of the proposed algorithm. Making decisions about the size of patches is a crucial step in this algorithm. Size of the patches should be determined after investigating the mage mask in such a way that no unknown patch will remain.
In the current study, some novel ideas about reducing the need to meticulously inspecting the image mask and making a decision about the size of patches will be discussed. One of these ideas is to select the missing patch, inpainting, and embedding in the damaged image, and then selecting the next missing patch from the inpainted image and completing the inpainting process. Another idea here is the weighted mean of the selected and inpainted patches that have been yielded by overlapping from the retrieved values in the damaged patches and the original image. Notably, the weight of each retrieved pixel is equal to the ratio of the number of undamaged pixels to the size of the patch they are placed in. The performance of proposed ideas is tested for some of the fundus images from the DRIVE public database, employing different factors. Also, a comparison with other methods using visually inspecting and quantitive criteria is carried out. Our findings suggest that the use of proposed ideas outperform diffusion-based inpainting and other sparse-based algorithm that didn’t use this idea.
