RESwinT: enhanced pollen image classification with parallel window transformer and coordinate attention

Pollen image classification is crucial for understanding allergic reactions and environmental impacts. In this study, we propose RESwinT, an enhanced deep learning model specifically designed for pollen image classification. RESwinT incorporates a parallel window transformer block with contextual information aggregation to expand the receptive field and facilitate information exchange between image patches. Additionally, a coordinate attention module is integrated to emphasize channel-specific features, improving the model’s focus on salient pollen characteristics. Experiments conducted on a locally developed dataset of eight allergenic pollen types from Beijing, China, demonstrate that RESwinT achieves state-of-the-art performance, with an F1-score of 0.985 and an accuracy of 98.58%, surpassing existing CNN and transformer-based methods. These results highlight the effectiveness of RESwinT in pollen image classification and its potential for wider applications in environmental monitoring and healthcare.