Framework
The framework of this study is shown in Figure 1. The salamander heads
in the images were automatically detected using YOLOv5 (Redmon et al.
2016) and used as either training or test images. The training data used
5 native and 5 hybrid individuals selected randomly from the two groups
(Table 1; Supplementary Figures 1 and 2). The 6 remaining native
individuals and 15 remaining hybrid individuals not used in training
were selected as test images. These images were resized to 224 x 224
pixels to ensure consistency in size. Additionally, augmentation
processing (rotation, crop, brightness, gaussian noise, color jitter,
and saturation) was added to the training data to prevent over-fitting.
After augmentation, 70% of the images for training and 30% for
validation were randomly separated for analysis.