An improved Swin transformer for sandstone micro-image classification

Facing the challenge of accurately identifying and classifying sandstone microscopic images in geological sciences, especially in regard to their critical role in stratigraphic analysis and resource assessment, an improved framework based on the Swin Transformer has been designed. Highly complex and detail-rich images which pose significant processing constraints are effectively handled by this approach. Initially, the framework incorporates a Spatially Adaptive Enhancement Module (SAEM) to improve detail capture and classification accuracy. Following this, the deployment of Local Perception Blocks (LPB) further refines our understanding of spatial information in images, facilitating the recognition of intricate textures. To tackle class imbalance, a Balance Adaptive Mechanism (BAM) is introduced, significantly boosting generalization capabilities. Through extensive experimental validation, this methodology has proven to significantly surpass existing models in sandstone microscopic image classification, highlighting its effectiveness and innovative edge in challenges of complex image recognition challenges. This framework, integrating cutting-edge attention mechanisms with advanced deep learning technologies and optimization strategies, presents a robust solution for recognizing sandstone microscopic images, among other complex image processing endeavors.