Efficient Reconstruction of Masked Image Embeddings for Region-Specific Satellite Image Retrieval
Abstract
How can we efficiently retrieve historical meteorological satellite images that are partially similar to a given query image? With advancements in satellite technology, it has become possible to acquire large volumes of high-resolution meteorological images. To efficiently retrieve past satellite images with similar weather patterns, image embeddings are commonly used in place of raw image data. However, conventional retrieval methods struggle when similarity needs to be assessed within a specific region of the image rather than the whole. A typical approach, masking and re-encoding every candidate image per query, is computationally expensive and impractical at scale. In this paper, we propose FRAME (Focus Region Aware Masking Embedding), a mask-aware embedding method that enables efficient region-based similarity search without re-encoding original images. FRAME solves the Masked Image Embedding Reconstruction (MIER) problem by applying a learned projection of the mask directly to the precomputed image embeddings. Experimental results on COMS satellite images demonstrate that FRAME reduces computation time by approximately 92% and FLOPs by over 99% compared to the baseline. Moreover, FRAME maintains high retrieval performance, achieving above 99% cosine similarity to the ground-truth masked embeddings and over 80% recall@1 in masked image retrieval tasks.