Research On The Model Of Generating Chinese Poems From Images Based On Neural Network

With the advent of the big data era and the introduction and application of various machine learning algorithms,especially the development of deep learning technologies,the design ideas of artificial intelligence have been promoted.Thanks to deep learning techniques,machines are increasingly like humans,and computers begin to understand higher level image content and text sequences as humans do.As the core technology of deep learning,deep neural network has made new breakthroughs in the field of automatically describing an image and automatic generation of Chinese ancient poetry due to its powerful feature learning and modeling capabilities.However,in the automatic generation of image description,the content of the image is described by a simple vernacular sentence,which is monotonous and inflexible.And the current Chinese poetry generation methods based on the designation of thematic words have great limitations.These methods have a high requirement for the selection of thematic words,and only when the selection of thematic words is reasonable,the generated ancient poetry is more reasonable,which will cause obstacles for many ordinary users.Moreover,these approaches are constrained by strict rules and patterns,lacking coherence and flexibility between sentences.To solve the above problems,this paper proposes a generative model based on deep neural network to implement the function of describing the image in the form of Chinese poems.The generative model can automatically convert images into Chinese ancient poems,getting rid of the restrictions of the thematic words,which allows ordinary users to enter an image to obtain Chinese ancient poems that describe the image.Our model consists of two parts,one is to extract information according to the semantics presented in images by a multimodal recurrent neural networks,and the other is to generate each line of the poem incrementally according to the extracted semantic information from the images by a recurrent neural network.Experimental results demonstrate the effectiveness of our approach by manual evaluation.