Research On Image Caption Quality Evaluation Based On Text And Semantics

In recent years,with the rapid development of artificial intelligence technology,natural language processing and computer vision have made tremendous achievements in understanding images and generating text,and have a wide range of practical applications.The machine's understanding and recognition of images are inseparable from image annotation data.There are two types of existing image annotation data:manual annotation and machine annotation.Because machine learning requires high quality and quantity of image annotation data,training efficient machine image annotation models and evaluating the quality of image annotation data are two important research directions in the field of image caption.At present,commonly used image caption quality evaluation methods are essentially based on the text level.Algorithms like BLEU and CIDEr all use text matching between machine captions and manual captions to evaluate caption quality.This leads to the lack of matching relationship between the image and the corresponding image caption,plus the ambiguity of natural language itself,which makes the standard of image caption quality assessment even more doubtful.In response to the above problems,this paper proposes a machine learning-based image caption quality evaluation framework RCWS(Region Rank Similarity-Consensus-based Image Description Evaluation and Weight Distribution Similarity-Semantic Propositional Image Caption Evaluation,RCWS),which not only considers images the relationship between itself and the content of the corresponding machine caption also considers the degree of matching between the machine caption and the manual caption.The basic idea is to use artificial image caption data as the standard and machine image caption data as the research object.By analyzing and combing the existing image caption data quality evaluation methods,the regional ranking similarity and weight distribution similarity are applied to image caption In the quality evaluation,the text-based image caption quality evaluation algorithm CIDEr and the semantic-based image caption quality evaluation method SPICE were improved.In this paper,the quality of machine image caption data is evaluated based onartificial image caption data.Through the three data sets of MSCOCO,Flicker 30 k and Flickr 8k,the usability and effectiveness of the proposed RCSW quality evaluation framework,evaluation model and improved image caption quality evaluation algorithms R-CIDEr and W-SPICE were verified.Experimental results show that the improved R-CIDEr and W-SPICE algorithms improve the evaluation performance of multiple data sets,and are superior to traditional evaluation algorithms in terms of text / semantic consistency.This achievement has good reference value and practical application value for artificial intelligence recognition,image caption quality evaluation and training of machine image caption models.