Visually impaired people have to rely on alt texts attached to images on web pages to complete their online tasks. However, too many alt texts of the unrelated images may increase the difficulty of reading. This study is intended to automatically filter important images of the web page by means of deep learning. After screening important images will automatically increase the image interpretation, so that the content of the web page is more easily understood by the visually impaired. This paper proposes a method for automatically filter important images on a web page. The research method is divided into four steps. First, generating image captions. Second, comparing text similarity with image titles and main articles. Third, choosing images with high similarity. Finally, the results of the automatic and manual methods are compared. In this paper, we tested our automated method and compared with the mutual method. tested 5 webpages includes with 1 shopping website and 4 news websites. The results are as follows. First, for the automated method, it is very hard to distinguish the similar images for different purpose. In the future, we will add other attributes like size, position to filter of important images. Second, there are some main images in the article is not high related to the article. In people’s thoughts, these images are important images, but it may be no need to add alt text at all. Third, in the automated method, accuracy of image caption and text length will affect the result.