Enhanced Pollen Classification via Hybrid Neural Network With Attention Mechanism and View Separation: An Equatorial and Polar Approach

Accurate classification of pollen grains is crucial for various fields, including ecology and food engineering. However, current approaches often face challenges due to limited data and the morphological variability of pollen grains. This study proposes a novel methodology emphasizing the separation of pollen images into equatorial and polar views, resulting in a significant improvement in classification accuracy. Using pseudo-labeling techniques, the dataset was segmented into these views, enabling the creation of customized datasets for classification. Experiments conducted with the Cretan Pollen Dataset, using a hybrid neural network combining of VGG19 and the CBAM attention mechanism, trained from scratch with data augmentation, demonstrated remarkable performance compared to previous studies. The VGG19+CBAM network achieved an accuracy of 98.78% and a precision of 98.72% with the dataset in its original format, and a precision of 98.06% with the view-separated dataset. This work highlights the fundamental role of view separation in enhancing classification results and advancing automated pollen analysis.