Research On Aesthetic Classification And Emotion Labeling Of Fabric Images Based On Convolution Neural Network

With the continuous improvement of people's quality of life,consumers not only consider their clothing as necessities,but also expand the functions of apparel products to social,spiritual and cultural fields.As an important fashion apparel,necktie can reflect owner's social identity,character,cultural taste and aesthetic taste.Different necktie patterns have different emotional expression.With the rapid development of clothing digital design,manufacturing and e-commerce,tie pattern samples are increasing day by day,thousands of patterns are stored in the sample library,it is difficult for designers to make full use of them.Simultaneously,consumers need to pick the right product quickly when shopping online.Aesthetic evaluation of tie patterns can effectively improve search performance based on semantic.This will not only help designers take full advantage of existing samples and inspire creative inspiration,but also help users select products meeting needs.Therefore,research on fabric aesthetic classification and emotion labeling can provide theoretical basis for the application of computer graphics and intelligent computing models in design,and have far-reaching great practical significance for clothing and fabric design.Nowadays,with the rise of deep learning and big data,some researchers try to classify images by needs of human emotions.The traditional methods based on manual features ignore the influence of image emotions and cannot meet the individual needs of designers and consumers.Aesthetic classification and emotional labeling of fabric patterns is the basis of fabric emotional retrieval.This thesis takes tie patterns as the research object,combines the characteristics of handmade aesthetics,presents an aesthetic classification method and an emotion labeling method based on parallel convolution neural network.The main work of this thesis is summarized as follows:(1)High and low aesthetic classification based on parallel convolutional neural network.First,necktie pattern images have been aesthetic evaluated subjectively,in order to establish a high and low aesthetic image sample library.Secondly,from the perspective of image aesthetics,handmade aesthetic features have been extracted such as Image edge,texture,saliency area and so on.Thirdly,the structure of convolutional neural network(number of network layers and filters)has been improved through experiment results,and the parallel convolutional neural network structure has been applied to high and low aesthetic classification.Then,the manual aesthetics and image pixel values have been input into the parallel CNN for training.At last,the CNN after training has been used to classify the high and low aesthetics of test sample images.The experimental results show that the accuracy of parallel CNN is higher than traditional manual methods,and higher than current popular convolution neural network models.(2)Emotional labeling method based on parallel convolution neural network.First,according to human emotion,costume theory,color psychology and other multidisciplinary theoretical research,five pairs of descriptive words that can describe fabric emotion have been synthetically selected,which are "formal"-"casual","gorgeous"-"elegant","complex "-"concise","classic"-"modern","active"-"boring".Numerous clothing field experts labeled emotion of images,established a sample library.Secondly,handmade emotion features such as image saturation and texture have been extracted from the perspective of image emotion.Thirdly,the structure of convolutional neural network(number of network layers and filters)has been improved through experiment results,and the parallel convolutional neural network structure has been applied to emotional labeling.Then,the manual aesthetics and image pixel values have been input into the parallel CNN for training.At last,the CNN after training has been used to label necktie patterns using emotions.The experimental results show that the accuracy of parallel CNN is higher than traditional manual methods,and higher than current popular convolution neural network models.The research in this thesis provides a new way of necktie emotion labeling,which can be used to assist the design and purchase of necktie.