Pollen Grains Image Data Augmentation and Processing Based on Unpaired Translation

Deep learning algorithms for pollen grains classification help monitor airborne pollen grains and forecast the risks of allergic reactions but are dependent on sufficient and balanced image datasets. Pollen grain image datasets are often imbalanced and deficient. This paper provides a series of methods of pollen grain data augmentation by performing feature enhancement/attenuation and staining normalization using unpaired translation. The BPDD-LM dataset was divided into different domains based on staining effects or visual features as inputs to train the model. The output images were evaluated by several metrics and a classification neural network. The results indicated that output images have the desired transformation and can be used to ameliorate the sufficiency and balance of datasets. The methods can be combined to increase the diversity of output and still remain effective for data augmentation.