Sports image retrieval based on hand-drawn sketches

Sketch-based sports retrieval provides an intuitive and efficient interactive way for sports video analysis and tactical research. This problem is very challenging, on the one hand, in how to group annotations and, on the other hand, in how to perform cross-modal alignment between sketches and real image shapes. In addition, sketches are usually highly abstract, resulting in weak semantic information expression ability, further affecting the retrieval effect. To address the above problems, we first constructed the largest sports sketch training data to date, including multiple sports such as basketball and football. Secondly, in order to effectively bridge the gap between sketches and real images, we proposed a cross-modal sports retrieval framework based on heterogeneous networks to effectively bridge the visual gap between sketches and videos. In addition, we introduced label word vectors to enhance the feature expression ability of sketches using semantic information and improve the semantic consistency of cross-modal matching. Finally, we designed a joint loss function for similarity learning, which improves retrieval accuracy and enhances model generalization ability by maximizing inter-class similarity and minimizing intra-class similarity. Experimental results show that the proposed method outperforms existing methods in terms of retrieval accuracy and computational efficiency, providing a new intelligent solution for sports analysis.