Research On Automatic Recognition Method Of Image Color Vocabulary Based On Deep Neural Network

Human color vision is a subjective perception,there may be differences in color perception between different individuals.Although there is a difference in individual color perception,it does not prevent people from discussing colors by using color vocabulary.At present,color displays control,color image storage and other fields in computer screen mainly using RGB value or CIE chromaticity coordinate value.However,the color corresponding to the RGB value is highly correlated with the display device,and the same RGB value will appear in different colors on different display devices.Even if they are physically identical color blocks(with identical CIE chromaticity coordinate values),there are different color perceptions in different spatial contexts.For example,in a color illusion image,the two color blocks are physically identical,and because of the different environmental effects,human color perception perceives two completely different colors.In view of the above problems,and considering that in daily life we all use color vocabulary to communicate,this paper first use the visual psychological physics experiment to establish a color label dataset based on human subjective color visual perception,and then use this dataset to train the deep neural network to achieve automatic recognition of the image corresponding to the 11 color vocabulary,and to get a output of 11 color labels at a time.You can set the number of output color labels,a color recognition method based on subjective visual perception can avoid the perceived differences caused by other color extraction methods.In this paper,several convolution neural networks with different structures are constructed as multi-label classifiers,and the training sets are divided into two categories,the first includes color block pictures and specific types of pictures,and the second category includes only specific types of pictures.Through the experiment,the following three kinds of comparison are made: two network structure comparison,training dataset comparison,threshold comparison,so as to determine the best recognition rate of color vocabulary recognition model,which can accurately identify a variety of color vocabulary,its evaluation index F1-measure in different test sets can reach more than 75%,Compared with other color quantization methods,it has higher accuracy and robustness in identifying the main color of the subject.In this paper,the design and establishment of the early color lexical data was set to obtain a higher quality data set to ensure the training effect of neural network,and the comparison of the construction of network structure,training set types and different thresholds is to ensure the recognition rate of color lexical recognition model.The model constructed in this paper solves some problems of the existing color extraction methods,such as the difference between the extracted color value and the visual subjective perception.The color vocabulary which is being identified is not only helpful to the image retrieval technology based on color clues,but is also more conducive to the description of color in our daily communication.