Chinese Image Description Generation Model Based on Recurrent Fusion Encoding

The image description job is a crucial area of research in the nexus of computer vision and applications for natural language processing. The objective is to produce accurate text descriptions using the picture attributes retrieved by the model. The conventional encoder-decoder model produces an inadequate encoding feature because it does less direct dimensionality reduction on images and is deficient in sequence feature information. This research suggests a recurrent fusion encoding-based Chinese picture description generation model to solve the aforementioned issues. R-CNN is used to extract the image's features first. Then, at the encoding side, gated recurrent units and a soft attention mechanism are used to build a recurrent fusion network for full sequence fusion of the image features. Finally, the sequence feature information is combined with a self-attention mechanism for multi-level interaction before being sent to the encoding side. I carried out enough tests on the Flickr_8k dataset and the Ai_Challenger dataset to confirm the model’s validity. According to the experimental data, the model performs better than some Chinese image description algorithms in each index by 0.2–4.2% points, and the generated statements are more accurate and sufficient for image description.