Abstract
Purpose :
Precise image registration and montage are critical for high-resolution adaptive optics (AO) retinal image analysis but are challenged by rapid eye movement, especially the saccades during continuous image acquisition. We developed a substrip-based registration algorithm to improve image registration and achieve an automatic montage of AO retinal images.
Methods :
For a succession of retinal images recorded continuously during the human subject following a moving fixation target, the algorithm first estimates the translation between 2 consecutive frames by fast Fourier transform (FFT) based phase correlation and divides the video into a series of image batches. Each contained successive frames with frame-to-frame translations smaller than a certain number of pixels. Next, the algorithm selects an image with minimal distortion as a reference frame in each group and divides each image into multiple strips (m × 512 pixels). Then, it calculates the normalized cross-correlation with the reference frame separately using 2 substrips (m × 64 pixels/each) at both ends of an image strip and takes the offset of the higher correlation peak as the relative translation of the whole strip. Next, the algorithm establishes a montaging coordinate system by connecting all group reference frames using the substrip-based registration and converting the frame-to-frame translations within each batch to this coordinate system, thereby generating a registered montage. We evaluated the algorithm with retinal images (512×512 pixels) acquired by an adaptive optics scanning laser ophthalmoscope (AOSLO).
Results :
The algorithm demonstrated robust registration ability. Any two frames with a motion amplitude of up to 448 pixels in the fast scanner direction can be precisely registered. In addition, automatic montage spanning up to 10 degrees on the retina was achieved on a cell-to-cell precision.
Conclusions :
The substrip-based method enables precise high-resolution AO image registration. Automatic montage reduces the manual labor required to generate high-quality AO large field-of-view retinal images and is promising to improve the productivity of image analysis.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.