combinations of flipping, 90° rotation, zoom, brightness, blurring, and noise, data augmentation was applied in Roboflow across four configurations: combined XPL-PPL images with threefold (3X) and tenfold (10X) augmentation, XPL images with tenfold (XPL-10X), and PPL images with tenfold augmentation (PPL-10X). The hyperparameters (Table 2) are training configuration values aimed at improving model performance and are the result of several previous training runs. Table 2 – Hyperparameters for the four image datasets: 3X, 10X, XPL-10X, and PPL-10X Hyperparameter Value NumWorkers; Batch Size; ROI Heads Batch Size 8; 8; 512 Optimizer ADAM Momentum 1; Momentum 2; Regularization L2 0.9; 0.999; 0.0001 Iterations 2000 Warm-up Phase (iterations) 500 Initial Learning Rate 0.05 Learning Rate Decay Factor 0.1 Learning Rate Step (iteration) 1250, 1750 Figure 4 shows the total loss for each training dataset. In the initial training, the model converged to a value of 2.3. After incorporating the optimizer, the loss decreased to 1.8. The inclusion of a training routine revealed that both values corresponded to local minima and that the applied adjustments improved model optimization. Regarding dataset performance, the XPL-10X model achieved the best results, followed by the 3X dataset. The 10X and PPL-10X datasets exhibit similar total loss values.
RkJQdWJsaXNoZXIy MTM0Mzk2