The Eyes are part of our body that are responsible for our visual inputs. Visual inputs are the first step to visual understanding that is a fundamental part of our survival and being. They are external organs that can be seen without any specialized tools that make them one of the first things you notice about a person. They are also unique to each individual whose colour can be different from person to person. The main eye colours are brown, black, blue, and green.
The coloured part of the eye is called the iris that has unique patterns for each person. Therefore, it can be used as a personal identity. The white, veiny part of the eye is called the sclera. The sclera is responsible for eye structure and protecting the eye from any harm. The small inner circle inside of the iris is called the pupil. The pupil works like a regulator that adjusts how much light can enter the eye. With the information they contain such as iris patterns and the unique vein structure of the sclera, they can be used to identify a human. Also, they show signs about the wellness of a human.
For example, redness of the sclera can be a sign of allergies or infections, the pupil might lose its brightness due to the cataract disease. They can lose their shape due to genetics or accidents. One of the ways to obtain eye information is called eye segmentation. Segmentation is the task in which input images are annotated to parts referring to which objects are in that image. Eye segmentation is the task used to get boundaries of the parts we want to use. For example, sometimes only iris is required for the research. Therefore, the eye image is annotated to have two parts iris and background.
What is The most common use case of eye segmentation?.
The most common use case of eye segmentation is iris recognition. Iris recognition systems use special hardware to capture eye image that is highly restricted and uses image processing methods to locate iris location. Main methods used for such task is called Hough Circle Transformation and Daughman’s Method. The main difference between those algorithms is Daughman’s Method does not expect the iris to be perfectly circular. Therefore, it is more commonly used in iris recognition task.
The problem with image processing methods is the expected input of repeated patterns, that are near impossible to get when we consider 3d model of the eye. Therefore, there is a need for a model that adapts different patterns in inputs. This is where neural networks excel. Segmentation using neural networks is a common practice since it’s being used in various problems such as segmenting the roads for autonomous driving cars, detecting faulty cells for disease detection, and even art style transformation.
Neural networks require input, output pairs to understand the problem and create the required filters to address a solution for the problem. For different images to be segmented, an adequate amount of manually segmented input and output pairs need to be given to train the network. Thus, this creates a challenge compared to image processing methods which only require certain calculations to do the task.
In order to satisfy training and testing data required for deep neural networks, several datasets are created. The first dataset contains images taken for BAP project TTU-2018- 3295. The second dataset contains 900 synthetic images generated from the UnityEyes interface. Alongside these datasets, 415 images from the MicheII challenge are also annotated. Datasets contain images are in different angles, distances, and lighting conditions. Thus, that makes them a good candidate for comprehensive image segmentation. In addition to sparse images, the effect of the data augmentation methods is also tested. The model used for training also affects the results.