Figure 2.

Overview of the adapted CNN used for the extraction of stomata morphometry. The proposed CNN combines features of both an Attention U-Net (Oktay et al., 2018) and Inception (Szegedy et al., 2016) to make pixel-level predictions of stomata for both guard cell and pore. The CNN contains a number of layers including convolution (Conv) layers, Max pooling layers, and fully connected layers. The output of each convolution layer is a set of 2D images, known as feature maps, which are computed by convolving previous feature maps with a filter, the size of which is given in the key. Batch normalisation (BN) and Rectified Linear Units (ReLU) steps are added to normalise data and remove negative pixel values from features maps. Skip connections help to maintain spatial information whilst the Attention Gate removes redundant features. The number of filters at each step is given as the blue number, whilst the resolution is given in black.