Thangka image caption method based on attention mechanism and encoder-decoder architecture

Abstract

The task of image description is to automatically generate sentences describing the image based on the input image, which belongs to the intersection of computer vision and natural language processing. Aiming at the characteristics of Thangka image with complex background and numerous objects, this paper proposes a Thangka image description method based on attention mechanism and encoder-decoder architecture. To begin, a convolutional neural network encoder is built, and an attention mechanism is added to it to better extract the features of thangka images. Second, construct a decoder based on a long short-term memory network to collect semantic features between words inside a sentence and to understand the mapping relationship between picture features and word semantic features. On the Thangka dataset and the public dataset Flickr8K, the model is validated. The results of the studies show that the model considerably improves the accuracy of providing descriptive sentences.