FINE-GRAINED SKETCH-BASED IMAGE RETRIEVAL USING SPATIAL RELATIVE FEATURE OF IMAGES

Fine-grained sketch-based image retrieval (FG-SBIR) aims to use hand-drawn sketches as queries sample to perform instance-level retrieval in image databases. Due to problems such as offset and distortion, it is difficult for sketches to align with real images in space. A new FG-SBIR method based on spatial relative attention (SRA) is proposed. The coordinate map was used to extract spatial-aware features, and the spatial relative relationship was extracted based on the relative self-attention module. At the same time, a contrastive loss NT-Xent was introduced to obtain stronger feature representation. Experiments on two public datasets show that the method outperforms several state-of-the-art FG-SBIR methods on retrieval accuracy.