4 50 4. End-to-end optimization of prosthetic vision Figure 4.7: Receiver-operator curves for the semantic boundary prediction task in experiment 3. Our proposed end-to-end method is compared against existing approaches: Canny edge detection (Canny, 1986) and holistically nested edge detection (Xie & Tu, 2017). The specificity (1 - False Positive Rate), sensitivity and area under the curve (AUC) of the thresholded predictions are also provided inTable4.3. phosphenes, the encoder is equipped with a fully connected output layer on top of the architecture which is shown in Table 4.1. In contrast with convolutional layers, spatial information is lost in a fully connected artificial neural network layer. To preserve spatial coherence between phosphene encodings and the training images, we introduce a regularization term to the cost function, that drives the network to activate phosphenes that correspond with bright regions of the training images and vice-versa. This spatial regularization loss is calculated as the BCE-loss (seeEquation(4.5)) between the output of the encoder and the pixels in the training target, sampled at the location of the phosphene center. The model successfully converged to an optimal solution and the results of experiment 4 are displayed inFigure4.8. The reconstruction performance, reflected by the AUC score, was significantly different amongst the various phosphene resolutions (Figure4.8a). The overall reconstruction performance was lower compared to experiment 3 (Figure4.7). Int. Perc. Sem. MSE SSIM FSIM MSE SSIM FSIM Acc. Sens. Spec. Prec. AUC E2E 0.034 0.554 0.719 0.063 0.541 0.761 0.697 0.722 0.695 0.165 0.785 Canny 0.055 0.443 0.672 0.061 0.454 0.581 0.598 0.774 0.583 0.134 0.746 HED 0.056 0.454 0.708 0.059 0.458 0.588 0.757 0.571 0.772 0.173 0.724 Table 4.3: Performance metrics for experiment 3. MSE: mean squared error; FSIM: feature similarity index; SSIM: structural similarity index; Acc.: accuracy, defined as the proportion of correctly classified pixels; Sens.: sensitivity, defined as the proportion of boundary pixels that were correctly identified as such; Spec.: specificity, defined as the proportion of non-boundary pixels that were correctly identified as such. AUC: area under the receiver-operator curve. Significant highest performances are indicated in bold.
RkJQdWJsaXNoZXIy MTk4NDMw