18 Application of a U-Net Convolutional Neural Network to Ultrasonic Wavefield Measurements for Defect Characterization 175 Fig. 18.6 Learning rate distribution over training iterations Fig. 18.7 Training and validation loss over training iterations Fig. 18.8 Intersection over Union (IoU) metric During the early iterations, the loss is shown to rapidly spike up and down due to the increased learning rate, with an overall decreasing trend. During later iterations, the loss decreases more steadily and progresses toward a more global minimum as the learning rate is refined. The training process was continued until the loss leveled to an asymptotic value. After training of the CNN was completed, its performance on the validation set was evaluated by a classic object detection metric, Intersection over Union (IoU). For a semantic segmentation task, IoU is defined as the area of overlap between the predicted pixel classifications and the ground truth labels divided by the area of the union, averaged over all segmentation classes, as illustrated in Fig. 18.8. The IoU metric is the most commonly reported benchmark for comparison against previously published models. In order to benchmark the training performance of the CNN, the final IoU value of the model was compared with previously reported values of other image segmentation CNN models, such as RefineNet [35] and DeepLab [27]. These results are summarized in Table 18.2. The final IoU value of the wavefield CNN was 76.3%. It is noted that the DeepLab and RefineNet IoU scores are not directly comparable to the wavefield CNN, since they were trained on a different test set (PASCAL VOC 2012 test set). They are provided here, however, to give a general sense of the range of IoU values reported by state-of-the-art image
RkJQdWJsaXNoZXIy MTMzNzEzMQ==