Fig. 4: Construction of convolutional denoising autoencoder. | Nature Communications

Fig. 4: Construction of convolutional denoising autoencoder.

From: Denoising-autoencoder-facilitated MEMS computational spectrometer with enhanced resolution on a silicon photonic chip

Fig. 4

a Noise-oriented training scheme for the denoising autoencoder. b Make the noise dataset. c Architecture of the convolutional autoencoder. d–f Optimization of convolutional neural network (CNN) structural parameters based on reconstruction resolution, including the number of convolutional layers, the number of residual blocks, and the kernel size. Boxplots indicate median (middle line), 25th, 75th percentile (box), and 5th, 95th percentile (whiskers). g Evolution of the mean squared error (MSE) and the resolution in each generation, with increasing generations. The top and bottom bars represent the maximum and minimum values of resolution, respectively.

Back to article page