Research on image captioning using dilated convolution ResNet and attention mechanism

Image captioning, which refers to generating a textual description of the image content from a given image, has been recognized as a key problem in visual-to-linguistic tasks. In this work, we introduce dilated convolution to increase the perceptual field, which can better capture an image’s details and contextual information and extract richer image features. A sparse multilayer perceptron is introduced and combined with an attention mechanism to enhance the extraction of detailed features and attention to essential feature regions, thus improving the network’s expressive ability and feature selection. Furthermore, the residual squeeze-and-excitation module is added to help the model better understand the image content, thus improving the accuracy of the image captioning task. However, the main challenge is achieving high accuracy in capturing both local and global image features simultaneously while maintaining model efficiency and reducing computational costs. The experimental results on the Flickr8k and Flickr30k datasets show that our proposed method has improved the generation accuracy and diversity, which can better capture image features and improve the accuracy of generated captions. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2025.