Fine-Grained Sketch Retrieval for Multi Metric Feature Fusion with Attentional Mechanism

Sketch retrieval is one of the research topics in the fields of computer vision and image retrieval, aiming to retrieve the corresponding real images through simple drawn image sketches. The current sketch retrieval methods often use single degree variables to learn depth features, which cannot fully utilize the characteristics of the sketch itself, affecting the retrieval accuracy. Therefore, we propose a multi metric feature fusion network with attention mechanism (MFFAN) to solve this problem, which processes sketch features through attention mechanism and multi-dimensional feature fusion network (MDFN), and combines regularized multi metric triple loss (RMM-triple loss) control to control the direction of network updates, thereby improving retrieval accuracy. The purpose of incorporating attention mechanism and fusing HOG is to focus on the correlation between pixels and fully extract the feature information of the image. Calculate the distance between positive and negative samples and anchor samples through a mixed metric of regularization. The experimental results on the ShoeV2 sketch retrieval dataset show that the feature fusion network from multiple angles can extract effective features and improve the retrieval performance of sketch retrieval.