Image Semantics Understanding Of Chinese Based On Deep Learning

Image semantic understanding is a semantic interpretation and description of image content. It is based on image identification and is a comprehensive discipline of computer science, psychology and linguistics. The image semantic understanding is one of the most important high-level tasks and widely used in vision field, such as image retrieval, image labeling, image analysis and so on. In recent years, though the image semantic under-standing has achieved some breakthrough results, the accuracy and the integrity of content description are still relatively low. Because of the particularity of Chinese, Chinese-based image semantic understanding has not yet made a significant breakthrough. In view of these problems, this paper focuses on Chinese-based image semantic understanding and the main contents of this paper are:Firstly, the image feature extraction method is studied. The conventional method performs pre-processing of the image and the design of image features artificially, which are time-consuming. Different with the conventional method, the deep learning, which has been popular in recent years, is adopted to express the image features. Through comparing the model complexity and accuracy of many different deep neural networks, an improved GoogleNet deep neural network is designed in this paper for image feature extraction.Secondly, the extracted image features of the convolutional neural network are taken as the input of the long-short term memory network, which is used to establish the lan-guage model and generate the image annotations. The generated image annotates express the background as well as the task of a picture or animal behavior with a meaningful de-scription. In addition, the attention mechanism, which is popular in machine translation domain in recent years, has been added in the decoding model. Based on the fact that human beings can constantly correct the results in the process of image comprehension,this paper presents a decoding model based on the double-layer attention mechanism.In training data pre-processing, this paper first studies the Chinese word segmenta-tion method. Based on the analysis of the advantages and disadvantages of traditional Chinese word segmentation method, this paper introduces the deep learning based word segmentation method to carry out the entity recognition of Chinese sentences. Then, all Chinese vocabularies are encoded as feature vectors with equal length and serve as the input tags of the decoding model.Finally, the model for Chinese-based image semantic understanding is trained and tested on the Flickr 8k public Chinese marked data set and the verification results show that the double-layer attention decoding model has achieved better accuracy in Chinese annotation than single-layer attention model and the decoding model without attention mechanism.