Multimodal machine translation can improve the accuracy and fluency of machine translation-generated translations, while overcoming the ambiguity problem that exists in traditional text-only machine translation tasks. In this paper, after preprocessing the original text information, the visual information features of the image are extracted using contextual information and convolutional neural network, and then the visual information features are deeply interacted and jointly encoded with the text information features. The text information and visual information can be more closely and accurately fused, so as to improve the comprehension ability of the English multimodal machine translation method. The experimental results show that the English multimodal machine translation method fusing visual information proposed in this paper can alleviate the problem of insufficient resources for real-time translation tasks with fewer samples by virtue of its own good multimodal comprehension ability, and the BLUE score of the model in this paper is improved by 1.04 compared with that of Transformer. It also improves the focus on noun-verb and acquires more semantic features. At the same time, the application of multimodal machine translation also provides richer data support for teaching assessment, which is conducive to the construction of more scientific assessment standards.