Development of a modified differential evolution for the problem of finding the pupil region of interest in real time

• Main Author: Grushko, Y. V.
• Format: Conference Proceeding
• Language: English
• Subjects: Algorithms; Degeneration; Differential equations; Evolutionary computation; Health care facilities; Hough transformation; Mutation; Numerical methods; Ophthalmology; Optimization; People with disabilities; Solution space
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0104019

Video oculography (VOG) occupies a key role in many theoretical and practical studies: in ophthalmology, biometric identification systems, in the task of augmentative communication for people with neuroparalytic syndromes. The pupil registration step of VOG process takes considerable time and computational resources in existing algorithms (Hough transform, Starburst, morphological erosion) or has limitations in accuracy. It is necessary to develop a technique that improves the quality of the VOG process in conditions of limited computing resources in order to reduce the cost of the technology for individual use by people with disabilities and medical institutions. The author proposes to present the process of pupil registration as a global multidimensional optimization problem and its solution by the stochastic method of differential evolution (DE). The optimization problem is formalized and efficiency of DE application in video oculography is shown. The modification of the numerical method of DE is proposed, including new models of mutation, crossover, selection. The modification is based on the process of formation genetic isolations in the neighborhood of all local and global extremums, and the subsequent growth of the most adapted isolation and degeneration of others, according to the differential equation of Verhulst-Pearl. The proposed method uses low values of initial population size (∼24 pcs. for VOG) and find fast (∼0.009 s. for ARM Cortex-A52) and accurate extremum in complex multimodal functions with small solution space. Dependencies of the results on the input parameters, for optimal tuning of the method to VOG, are given.