Plant Phenomics / 2021 / Article / Fig 3

Research Article

Automatic Fruit Morphology Phenome and Genetic Analysis: An Application in the Octoploid Strawberry

Figure 3

Autoencoder architectures: (a) architecture of convolutional autoencoder applied to the internal fruit images; (b) architecture of convolutional variational autoencoder applied to external fruit. Unlike classical autoencoders, variational autoencoders are generative process as they learn the parameters of a distribution, instead of the feature representation. The last network was trained using an image of size , the encoder step consisted on 4 convolutional layers with a kernel size equal to 3, and the linear rectified “ReLU” as activation function to perform feature extraction (see details in GitHub account). Finally, the convolution output is flattened, and the mean and sigma parameters are extracted from a dense layer. In the last network, the decoder step starts with a vector sampled from the latent distribution as input and reconstructs the input by performing deconvolution operations. The last deconvolution uses the sigmoid as activation function. The loss function is the Kullback-Leibler (KL) divergence, which consists of both a “reconstruction” and a “regularization” term. The first network is a classical autoencoder, which uses the classical mean squared error (MSE) as loss function.