Deep Convolutional Generative Adversarial Network-Based EMG Data Enhancement for Hand Motion Classification

Figures

• 

  FIGURE 1. It briefly describes the methods used in this article including matrix transformation, normalization, histogram equalization, and DCGAN.

• 

  FIGURE 2. It shows the grayscale images before and after equalization. The top two images are synthetic images, and the bottom two are real ones.

• 

  FIGURE 3. (A) is the construction of the generator, (B) is the construction of the discriminator.

• 

  FIGURE 4. The first, second, and third rows show the synthetic images with sample numbers of 5000, 200, and 80, respectively. The first five columns show the synthetic images at different epochs. The last column shows the real images.

• 

  FIGURE 5. It shows the EMG synthetic images trained by 80 samples at different epochs and compares synthetic images at epoch 1000 with the real image.

• 

  FIGURE 6. It is a conceptual diagram of the experimental setup and data collection scenario. It describes the process of collecting EMG signals of 13 gestures under 16 channels.

• 

  FIGURE 7. Experimental comparisons under different classifications, in which (A), (B), and (C) correspond to five-, three-, and two-class classification problems, respectively. Each bar represents the classification accuracy, and the error bar represents the corresponding standard deviation. Real100 means cross-validation with 100 real images for each gesture. Fake 100 represents cross-validation with 100 fake images for each gesture. Fake100+Real100 means cross-validation with 100 fake images and 100 real images mixed for each gesture.

• 

  FIGURE 8. (A) compares the cross-validation accuracy across different datasets by SVM-based classification, and (B) compares the cross-validation accuracy across different datasets by RF-based classification. R and F represent that only real data and fake data are used, respectively. R-E and F-E represent the real and fake datasets after equalization, respectively. FR represents a mixture dataset with equal amounts of real data and fake data. FR-E represents a mixture of equalized datasets containing both real and fake data.

• 

  FIGURE 9. 10-fold cross-validation accuracy with different datasets, including real data (R), fake data (F), real data after equalization (R-E), fake data after equalization (F-E), the mixture of fake data and real data (FR), the mixture of fake data and real data after equalization (FR-E).