Abstract
Purpose :
Correcting presbyopia by using optoelectronic focus-tunable lenses may provide a near-natural visual perception across distances. A current limitation of such devices is the lack of accurate and instant estimates of the subject gaze distance to apply the corresponding correction. In this context, we explored how neural networks to process eye images can be used to dynamically drive the spectacles.
Methods :
A customized device with two optoelectronic lenses and dynamic control has been developed. It is completely wireless and includes two cameras, which constantly record images from the subjects’ eyes. Those images are sent to a smartphone which forwards them to a computer where the inference is performed. The computer quickly returns the calculated gaze distance and the corresponding correction is applied to the optoelectronic lenses for correction. To train the neural network (NN) several images were alternatively displayed at three screens placed at 2.85m, 1m and 0.33m. The collected data includes the images of the subjects’ eyes and the distances where they were looking at. Blinks are automatically removed by the eye tracking algorithm. The initial dataset included 108,817 images from 6 different subjects, and a total of 33 recordings were performed. The used NN is based on a Resnet34 architecture, which is a DNN-based novel approach to difficult image classification and regression problems.
Results :
To achieve good training, we divided the data into three different groups. A complete recording (4,387 frames) was isolated for final evaluation of the neuronal network. Finally, the remaining data was split into two different sets, training (83,550 frames) and validation (20,880 frames). After completing the training, the average error obtained in the evaluation dataset was 0.02 D. However, the evaluation set is composed of frames from the same recordings used for training. To provide a better understanding of the accuracy of the DNN, a complete recording that the network had never seen was used for testing. The results presented only 0.53% frames with an error larger than 0.25 D.
Conclusions :
The use of neural networks for estimating gaze distance is a promising approach that enables automatic and calibration-less presbyopia correcting spectacles. This approach could be extended and applied beyond our own prototype and used for distance estimation in general-use eye-tracker devices.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.