Pathological Myopia Image Recognition Strategy Based on Data Augmentation and Model Fusion

Table 1

12 DAMFs.

No.	Training set name	DA method	Quantity
1	PALM-Training800-overturning	Original dataset + random flip (4 directions: up, down, left, and right)	800
2	PALM-Training800-noise	Original dataset + Gaussian white noise	800
3	PALM-Training800-color	Original dataset + randomly changing colors (brightness, contrast, saturation)	800
4	PALM-Training800-cropping	Original dataset + random cropping	800
5	PALM-Training800-deforming	Original dataset + random scaling, stretching (stretched into a square by the length or width of the images)	800
6	PALM-Training800-dimming	Original dataset + change clarity	800
7	PALM-Training1600-overturning-noise-color	Randomly stack method 3 or 4 (serial number) on the basis of PALM-Training800-overturning	1600
8	PALM-Training1600-overturning-cropping-deforming	Randomly stack method 5 or 6 (serial number) on the basis of PALM-Training800-overturning	1600
9	PALM-Training1600-overturning-dimming	Randomly stack method 7 (serial number) on the basis of PALM-Training800-overturning	1600
10	PALM-Training3200-overturning-noise-color-cropping-deforming-dimming	Randomly superimpose method 5 or 6 or 7 (serial number) on the basis of PALM-Training800-overturning-noise-color	3200
11	PALM-Training800-imgaug1	Original dataset + random cropping with 0–50 pixels around, 50% probability horizontal flip, Gaussian blur (sigma = 0 to 3.0)	800
12	PALM-Training1600-overturning-dimming-imgaug2	PALM-Training800-overturning-dimming dataset + multiple mixed random overlay	1600