Search

Plenoptic photography stands out in computational photography due to its ability to capture the light field by a single-shot single camera. It differs from conventional photography in the capacity to preserve the directional information of the light rays, in addition to capturing its intensity. This allows the formation of images with different perspectives and focus settings. The related literature exhibit several approaches to perform these processes however there is a lack of documented comparisons between them. This work presents the implementation of refocus algorithms for plenoptic images and proposes some strategies to compare their performance according to computation times, photometric features preservation and depth resolution.

Slit Lamp (SL) is the equipment commonly used in most eye clinics. This instrument allows to observe ocular structures of both the anterior segment and posterior segment with different values of magnification. Today it is very common, obtain images from SL with a Smartphone by placing a simple adapter in the eyepiece of it. While these images help documentation, they lose their diagnostic quality with this acquisition technique, mainly because of the lack of correction of optical aberrations and noise introduced in such a system. This paper analyzes the hardware settings and characterization of images needed to achieve a 2D image of diagnostic quality in SL. For this purpose was developed a universal adapter for SLR cameras and using the software capture the camera itself. They were captured 128 images: 73 of anterior segment, 25 fundus and the rest with a magnifying glass and panfunduscopio. These images were characterized by type and noise power. It was determined the point spread function PSF and the consequent MTF for the optical system of the full SL with all of its gains later managing to improve photographs by Lucy-Richardson deconvolution algorithm. Finally were obtained in the SL the numerical aperture and F number (#f). These latest data, uncommon in the technical data of manufacturers, are of fundamental importance in future adaptations that seek to make, preserving diagnostic quality and not just for documentation purposes.

This paper presents a workflow for automatic ex-traction of the vasculature in color retinal fundus images, optimized for the detection of smaller caliber vessels. This workflow is raised in three stages: preprocessing, processing and post processing. These stages combine various methodolo-gies of lighting correction, enhancement, detection and elimi-nation of spurious structures in the detected vessels. The methodologies selected from the state of the art were tested according to robustness, simplicity, reliability, velocity and affordability. The selection of methodologies was based on criteria of sensitivity, specificity, accuracy and positive and negative predictive values. The best performance was obtained by combining lines of work based on matched filter for detect-ing large and medium vessels caliber, whereas of smaller was obtained with binary morphology through modified Top Hat transform. The evaluation of the proposed workflow was per-formed using three public databases of fundus images and the Gold Standard. The sensitivity achieved for the DRIVE data base exceeds 73%, in STARE 82% and 72% ARIA, when the mask Gold Standard is formed by the segmentation of a single expert observer. These values increase if the mask Gold Stan-dard provides the dispersion between the segmentation of two experts: the sensitivity is increased to 86

We present a support system for changes detection in fundus images of the same patient taken at different time intervals. This process is useful for monitoring pathologies lasting for long periods of time, as are usually the ophthalmologic. We propose a flow of preprocessing, processing and postprocessing applied to a set of images selected from a public database, presenting pathological advances. A test interface was developed designed to select the images to be compared in order to apply the different methods developed and to display the results. We measure the system performance in terms of sensitivity, specificity and computation times. We have obtained good results, higher than 84% for the first two parameters and processing times lower than 3 seconds for 512x512 pixel images. For the specific case of detection of changes associated with bleeding, the system responds with sensitivity and specificity over 98%.

Diabetic retinopathy (DR) is a chronic disease of the ocular retina, which most of the times is only discovered when the disease is on an advanced stage and most of the damage is irreversible. For that reason, early diagnosis is paramount for avoiding the most severe consequences of the DR, of which complete blindness is not uncommon. Unsupervised or supervised image processing of retinal images emerges as a feasible tool for this diagnosis. The preprocessing stages are the key for any further assessment, since these images exhibit several defects, including non uniform illumination, sampling noise, uneven contrast due to pigmentation loss during sampling, and many others. Any feasible diagnosis system should work with images where these defects were compensated. In this work we analyze and test several correction techniques. Non uniform illumination is compensated using morphology and homomorphic filtering; uneven contrast is compensated using morphology and local enhancement. We tested our processing stages using Fuzzy C-Means, and local Hurst (self correlation) coefficient for unsupervised segmentation of the abnormal blood vessels. The results over a standard set of DR images are more than promising.