A joint intelligent retrieval method for multimodal data based on deep reinforcement learning algorithm

In order to address the issues of modal loss and singular fusion strategies in the field of multimodal entity alignment, a multimodal entity alignment method based on multi-level feature fusion and contrastive learning is proposed. Adopting a deep Q-network (DQN) intelligent retrieval framework, feedback such as click and dwell time is converted into environmental rewards to dynamically optimize retrieval weights. Build a cross modal embedding space and use contrastive learning to align feature representations of different modalities. The strategy network generates joint intelligent retrieval results for multimodal data. The experimental results show that this method significantly improves the accuracy and user satisfaction of multimodal data retrieval. Improved retrieval accuracy in scenarios such as intelligent security.