Multi-type Chinese Description Generation System For Images

Image description is a task that combines computer vision and natural language processing,transforming the input image directly into text that people can understand.Image description has strong practical value in image semantic assisted understanding and image retrieval.The generation of various types of descriptions of images has certain reference value in terms of ease of use,practicability,and innovation.This paper proposes a multi-type Chinese description generation method for images,and develops a set of image multi-type Chinese description generation system.Description types include Chinese text descriptions,ancient poetry and good sentences.For the generation of Chinese text description,based on the current English text description algorithm with better effect,a better image feature extraction layer is adopted.Due to the ambiguity of Chinese word segmentation,a word-based description generation method is adopted.Since the image features of different regions cannot be weighted during the text generation process,the attention mechanism is used to calculate the image feature weights of different regions at different times.Combining the optimal hyperparameters and large-scale Chinese text description data sets,an excellent Chinese text description model is trained.At the same time,the text description is used to extract keywords for generating ancient poetry and famous quotes.For the generation of ancient poetry,four key words in poetry are associated with keywords extracted from text description.Then the double-ended multi-layer LSTM structure is used to encode the keywords of the current sentence and the above information,and the coding based on attention mechanism is adopted.The encoder-decoder model generates the current verse.For the generation of famous sayings,by comparing and analyzing the results of the deep learning automatic generation method and the keyword matching method,it is determined that the keyword matching method is finally adopted,matching the most suitable sentence in the crawled sentence dataset.For the development of the system,a simple and easy-to-expand system was designed and developed,and integrated with the trained model.The entire system has strong availability,and the description generation time is basically controlled within three seconds.If further better computing resources and more complete marker data can be obtained,the algorithm described in this paper is expected to achieve greater breakthroughs in parameter models and evaluation indicators.