GDCSF: Global Depth Convolution-Based Swin Framework for Electron Microscopy Pollen Image Classification

Pollen allergies have emerged a seasonal epidemic, characterized by a high incidence rate, significantly impeding individuals’ pursuit of a healthy lifestyle. Pollen image classification holds significant importance for botanical research and environmental monitoring. However, in most cases, pollen classification still relies on manual and labor-intensive methods, making the process tedious and inefficient. Thus, the current greatest challenge lies in enhancing model accuracy and computational efficiency in handling highly complex and detail-rich images. We constructed a large and diverse dataset containing 125 types of pollen. This study proposes a new Global Depthwise Convolution Swin Framework (GDCSF) suitable for electron microscopy pollen image classification. The GDCSF model, composed of the Strip Convolution Scaling Module (SCSM), Field Dual Bilateral Attention Module (FDBA), and Swin Transformer module, aims to enhance the model’s comprehension and classification accuracy concerning intricate pollen structures. By integrating the SCSM module, the model effectively scales the input image dimensions, enabling it to capture long-distance contextual dependencies along the horizontal and vertical axes of the image. The FDBA module conducts field dual bilateral attention calculations to obtain channel features, addressing the deficiencies in extracting features from electron microscopy pollen channels and providing richer information for the GDCSF classification network, thereby further enhancing the model’s usability and efficiency. Compared to the established advanced Swin Transformer model, our experimental results demonstrate a significant improvement in accuracy achieved by our model, ultimately reaching a precision of 94.8%, outperforming most CNNs with better recognition performance and all experiments were trained from scratch.