Reference remote sensing image scene segmentation based on multi-scale cross-axis attention network

Segmentation of scenes in remote sensing images represents a crucial and demanding area in the processing of remote sensing image data, merging the fields of computer vision and natural language processing. Traditional image segmentation methods often struggle with the complex spatial scales of remote sensing images and the issue of fuzzy boundaries in segmentation targets, resulting in suboptimal outcomes. To address these challenges, we introduce the Multimodal Information Interaction Network (IMNet), a novel method specifically designed to meet the unique needs of scene segmentation in remote sensing imagery. First, we design an Intra-Modal Information Mining Module (IMM) to effectively handle fine-grained details at multiple scales. Additionally, to better capture global information, we introduce a mechanism to compute double cross-attention between two parallel axial attentions. Our experimental results indicate that our method attains a mean Intersection over Union (mIOU) of 80.46% on our dataset, outperforming current state-of-the-art models and establishing a new standard in the performance of remote sensing image scene segmentation.