IRNet: an Information Retrieval Model Based on Dual-Stream Fusion of CNN and Transformer

Convolutional neural network (CNN) and Transformer models are widely used in information retrieval. However, Transformer ignores local words for accurate retrieval and CNN lacks long-distance dependency modeling. To leverage the advantages of both, this paper constructs a CNN- Transformer dual-stream parallel network architecture. The model introduces location coding and local awareness mechanism to improve Transformer's input representation for better integration of local and global information. It also designs a new feature fusion module with an adaptive weight adjustment strategy to fuse features and preserve local and global features. Additionally, an improved attention mechanism is introduced to enhance key information extraction. Experiments show that the model improves recall score, average precision and frame rate by about 7%, 8% and 3% respectively compared to the traditional model, enhancing retrieval efficiency and accuracy.