A Bio-Inspired Visual Perception Transformer for Cross-Domain Semantic Segmentation of High-Resolution Remote Sensing Images

Abstract

Pixel-level classification of very-high-resolution images is a crucial yet challenging task in remote sensing. While transformers have demonstrated effectiveness in capturing dependencies, their tendency to partition images into patches may restrict their applicability to highly detailed remote sensing images. To extract latent contextual semantic information from high-resolution remote sensing images, we proposed a gaze–saccade transformer (GSV-Trans) with visual perceptual attention. GSV-Trans incorporates a visual perceptual attention (VPA) mechanism that dynamically allocates computational resources based on the semantic complexity of the image. The VPA mechanism includes both gaze attention and eye movement attention, enabling the model to focus on the most critical parts of the image and acquire competitive semantic information. Additionally, to capture contextual semantic information across different levels in the image, we designed an inter-layer short-term visual memory module with bidirectional affinity propagation to guide attention allocation. Furthermore, we introduced a dual-branch pseudo-label module (DBPL) that imposes pixel-level and category-level semantic constraints on both gaze and saccade branches. DBPL encourages the model to extract domain-invariant features and align semantic information across different domains in the feature space. Extensive experiments on multiple pixel-level classification benchmarks confirm the effectiveness and superiority of our method over the state of the art.