Hybrid-TransCD: A Hybrid Transformer Remote Sensing Image Change Detection Network via Token Aggregation

Existing optical remote sensing image change detection (CD) methods aim to learn an appropriate discriminate decision by analyzing the feature information of bitemporal images obtained at same place. However, complex scenes in high-resolution (HR) cause unsatisfied results, especially for some irregular and occluded objects. Although recent self-attention-driven models with CNN achieve promising effects, computational consumed parameters costs emerge as impassable gap HR images. In this paper, we utilize a transformer structure replacing self-attention stronger representations per image. addition, concurrent vision only consider tokenizing single-dimensional tokens, thus failing build multi-scale long-range interactions among features. Here, propose hybrid module detection, which fully representation attentions scales each via fine-grained mechanism. The key idea is establish heterogeneous semantic tokens containing multiple receptive fields, simultaneously preserving large object For building relationships between features without embedding token sequences from Siamese tokenizer, also introduced difference decoder (HDTD) layer further strengthen global dependencies high-level Compared capturing single-stream our HDTD directly focuses representing differential increasing exponential cost. Finally, cascade (CFD) aggregating different-dimensional upsampling establishing skip-connections. To evaluate effectiveness proposed method, experiments on two CD datasets are conducted. state-of-the-art methods, Hybrid-TransCD achieved superior performance both (i.e., LEVIR-CD, SYSU-CD) improvements 0.75% 1.98%, respectively.