DictCR-former: Content-aware dictionary transformer for cloud removal

Atmospheric conditions frequently result in variable cloud coverage in remote sensing imagery, posing significant challenges for downstream applications. While deep learning has demonstrated effectiveness in cloud removal, existing single-image approaches may not sufficiently compensate for critical information voids caused by cloud occlusion, making accurate restoration particularly challenging. To address these limitations, we propose DictCR-former, a transformer-based architecture that incorporates dictionary-inspired operators to enable cross-image feature aggregation through our Content-Aware Dictionary Attention (CADict-Att). The CADict-Att synergizes two core components: (1) Content-Aware Dictionary Learning (CADL) that dynamically constructs dictionary atoms by capturing latent cross-image correlations to encode fundamental semantic patterns; and (2) Dictionary Aggregation (DA) which adaptively combines contextually relevant atoms through learnable attention weights, reconstructing cloud-contaminated regions while maintaining spatial consistency with unobscured areas. Extensive experimental evaluations on multiple benchmarks demonstrate that DictCR-former outperforms existing methods in cloud removal, achieving superior performance while maintaining computational efficiency.