Fig 1.
Sample images for each pollen type.
Fig 2.
Convolutional neural network architecture and operation.
Image based on a similar figure published in [26].
Fig 3.
Cross-validation schematic for setups A and C.
Fig 4.
Cross-validation schematic for setup B.
Table 1.
Average correct classification rate over the test set for each setup.
Results from [6] are computed from their confusion matrices. In parentheses under the values, standard deviation.
Fig 5.
Correct classification rate for human operators and the best model reported by [6], together with the three proposed deep learning setups.
Fig 6.
Confusion matrix for the test set in setup A.
Fig 7.
Confusion matrix for the test set in setup B.
Fig 8.
Confusion matrix for the test set in setup C.
Table 2.
Measures for computational complexity.
Fig 9.
Distribution of the correct classification rate for the test set in setup C, by pollen type.
Types: A) Anadenanthera colubrina, B) Arecaceae, C) Arrabidaea, D) Cecropia pachystachya, E) Chromolaena laevigata, F) Combretum discolor, G) Croton urucurana, H) Dipteryx alata, I) Eucalyptus, J) Faramea, K) Hyptis, L) Mabea fistulifera, M) Matayba guianensis, N) Mimosa somnians, O) Myrcia, P) Protium heptaphyllum, Q) Qualea multiflora, R) Schinus terebinthifolius, S) Senegalia plumosa, T) Serjania laruotteana, U) Syagrus, V) Tridax procumbens, W) Urochloa decumbens.