Enhancing Electric Power Industry Image-Text Matching with Image Properties

The electric power industry has many valuable images containing meaningful information, such as on-site physical and technical schematic images, which can guide employees in operation and learning. However, these images sleep in documents because they are difficult to retrieve. The current image-text matching methods mainly rely on projecting the images and the text describing the pictures into shared embedding space, but they often encounter issues with imbalanced feature representations, which can compromise the reliability of retrieval outcomes. To tackle these challenges, we present an innovative Image-Text matching method that explicitly adopts image properties. Additionally, we proposed a new model structure to learn how to match visual content and textual descriptions. Experiment results on MSCOCO, Flickr30K, and the electric power industry images show that our proposed method brings improvements over existing methods. Further case studies show that image properties finally help obtain a more comprehensive representation.