Entropy Based CNN for Segmentation of Noisy Color Eye Images using Color, Texture and Brightness Contour Features

Abstract

Today’s most of the iris recognition systems are strongly dependent on user’s cooperation during image acquisition such as stop-stair condition, head position and camera distance. Images are taken in NIR spectrum to reduce the noise such as effect of illumination. Challenges faced by existing iris recognition systems are such as they are time consuming due to need of extra hardware setup and unable to achieve better performance for images acquired on-the-move, at-a-distance, etc.. To overcome these challenges, in this paper we proposed novel segmentation algorithm based on content based image retrieval approach. In proposed segmentation method, color, texture and brightness contour features were extracted. Entropy value for these extracted contour features was measured to reduce the dimensionality of features. These set of calculated entropy value was given as input to convolutional neural network to cluster noisy eye image into iris, sclera and pupil region. The proposed segmentation algorithm was experimented on freely available UBIRIS.V2 noisy eye image database using MATLAB. The experimentation results shows that proposed segmentation method is superior as compared to existing method by reducing average segmentation time up to 0.9sec and increasing segmentation accuracy up to 97% for non ideal color eye images.