Dataset supporting the publication "Pollen image manipulation using latent space"

Dataset supporting the publication "Pollen image manipulation using latent space" published in Frontiers in Plant Science

This dataset contains:

Figure 1. Block diagram concept of the study

Figure 2. Schematic of the StyleGAN neural network (formed of the mapping and synthesis networks) for generating images, and of then the use of a CNN for subsequent classification of the generated images.

Figure 3. Diagram of StyleGAN with 3 networks (Left) The mapping network transforms a random input into a style signal, controlling various aspects of image generation. (Middle) The synthesis (generator) network uses the information (A) from the mapping network to generate images from low to high resolution. It also incorporates random noise (B) to introduce variations and fine details. (Right) The discriminator network compares real and generated images, updating the weights of all networks through adversarial training to enhance performance.

Figure 4. Graph showing the accuracy of training and validation progress during training of the CNN.

Figure 5. Schematic of methodology of projecting an image into latent space, by generating random z vector, generating an image then comparing that image with the projected to obtain the suitable vector in latent space. The vector is then manipulated via adding or subtracting a vector before the synthesis network generates a new image.

Figure 6 Histogram of distribution of taxa in training dataset and generated dataset (as predicted by CNN).

Figure 7. Generated images of pollen grains created through latent w-space vector manipulation, showing the addition of a ‘size’ vector (-100%, -50%, 0%, +50%, +100%) added in the horizontal direction and a ‘spike’ vector (-100%, -50%, 0%, +50%, +100%) in the vertical direction, to generated images of (a) Knightia and (b) Coriaria. Each generated image also shows the predicted pollen taxa and predicted confidence, as well as the pollen size in pixels. Labelling is omitted from of images without any visible grains.

Figure 8. Generated images of pollen grains created through latent w-space vector manipulation, showing the addition of a ‘size’ vector (-100%, -50%, 0%, +50%, +100%) added in the horizontal direction and a ‘round’ vector (-100%, -50%, 0%, +50%, +100%) in the vertical direction, to generated images of (a) Metrosideros and (b) Disphyma. Each generated image also shows the predicted pollen taxa and predicted confidence, as well as the circularity of the pollen grain. Labelling is omitted from of images without any visible grains.

Figure 9. Generated images of pollen grains created through latent w-space vector manipulation, showing the interpolation of projected images between (a) Knightia and Kunzea (b) Brachyglottis repanda and Citrus. Each generated image also shows the predicted pollen taxa and predicted confidence.