An image caption generation method based on improved attention mechanism

Image captioning is a research hotspot in the field of image understanding. It spans the fields of computer vision and natural language processing. Most of the image captioning researches use deep learning technology, especially using Convolutional Neural Networks(CNN) as encoder-decoder structure for feature extraction. The existing captioning model is not of high quality, the important parts of the image are not sufficiently extracted features and the model is too complex. To address these problems, this paper introduces a new generation method based on attention mechanism. First, an efficient attention module is introduced to improve the feature extraction effect based on traditional attention mechanism. It reduces the model complexity and better extracts the important part of image features. The method uses convolutional features extracted from a pre-trained CNN model which integrate the features with the attention mechanism. Second, we also select the GRU language model as the decoder to construct descriptive sentences. Finally, experiments on the MSCOCO dataset and Flickr30k dataset show that the model proposed in this paper effectively improves the quality of descriptive sentences.