Fig. 6. Recognition accuracy of the hand-drawn hydrocarbon test set of trained neural network with different training/validation datasets. (a) Results of training with varying ratios of augmented and degraded hand-drawn hydrocarbon to synthetic data training sets (500 000 image total) and hand-drawn hydrocarbons validation set. (b) The effect of fine tuning is investigated by restarting the weights from training with a 500 000-image synthetic dataset used for both training and validation, and a 500 000-image synthetic dataset used for training with a hand-drawn validation set. The weights are restarted with a training set consisting of 90% synthetic data and 10% augmented and degraded hand-drawn data, and a validation set of hand-drawn hydrocarbons.