Deep-Learning Image Stabilization for Adaptive Optics Ophthalmoscopy

An adaptive optics scanning laser ophthalmoscope (AOSLO) has the characteristics of a high resolution and small field view (FOV), which are greatly affected by eye motion. Continual motion will cause distortions both within frame (intra-frame) between frames (inter-frame). Overcoming achieving image stabilization is first step great importance in analysis. Although cross-correlation-based methods enable registration to be achieved, manual identification distinguishing images with saccades required; accuracy, but it time-consuming complicated. Some imaging systems able compensate for during process, special hardware devices need integrated into system. In this paper, we proposed deep-learning-based algorithm automatic stabilization. The used VGG-16 network extract convolution features correlation filter detect position reference next frame, finally, compensated displacement achieve registration. According results, mean difference vertical horizontal was 0.07 pixels 0.16 pixels, respectively, 95% confidence interval (−3.26 px, 3.40 px) (−4.99 5.30 px). Pearson coefficients displacements these two were 0.99 0.99, respectively. Compared methods, had higher automatically removed blinks, corrected saccades. registration, enabled accuracy achieved without intervention.