Picture title generation method and device and computer readable storage medium

Abstract

The embodiment of the invention discloses a picture title generation method and device and a computer readable storage medium. A picture and corresponding document information are obtained from a database; an initial title corresponding to the picture is generated by using a coding and decoding model in a neural network; according to a preset matching rule, the matching degree between the initialtitle and statements in the corresponding document information is calculated to determine a title of the picture. The situation that the document information corresponding to the picture can more accurately reflect specific features of the picture is taken into account, and the picture is named in combination with the document information, so that the information of the obtained title is more abundant, the picture is described more accurately, and the situations that picture titles are too broad in the prior art and the repetition rate of the titles of the same type of pictures is high are effectively avoided.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method, device and computer-readable storage medium for generating image titles. Background technique [0002] Image caption generation is a comprehensive problem that combines computer vision and natural language processing. Generating a caption for a picture to describe the information contained in the picture is a research hotspot in recent years and has made substantial progress. At the same time, this is also a very challenging task. We not only need to determine which entities are contained in the picture, but also express the relationship between them and generate reasonable and fluent sentences. [0003] At present, the research methods of image caption generation can be roughly divided into the following two categories: one is a method that combines visual detectors and language modeling. This method first uses multi-instance learning to train a visua...

AI Technical Summary

Problems solved by technology

[0004] However, the methods of generating picture titles in the prior art are all based on the characteristics of the picture itself, and often contain the following defects: the generated picture titles are too broad to provide specific information, such as: "a tennis player is playing tennis", " A building", but we have no way of knowing who the athlete is and which building it is

Moreover, the discrimination of the generated titles is poor, that is, the probability of generating the same title for similar pictures is high. For example, whether it is Federer or Nadal, they can only generate "a tennis player is playing tennis".

Images