| The clothing pattern is a form of design that reflects the important information of the trend of clothing.In order to quickly automatically obtain clothing pattern information,improve the multi-characteristic recognition efficiency of clothing images,improve the problem of high-efficiency and low-efficiency characteristics of traditional clothing pattern design models,and finally promote the digital and intelligent development of the clothing industry.The paper proposes the use of deep learning recognition and generating clothing patterns to realize the combination of art and modern technology.The following is the specific research content:In terms of clothing pattern recognition,this article considers three elements: style,color,and pattern,and establishes one containing 3 types of clothes styles,6 colors,6patterns,a total of 15 types of sample libraries.The identification with the style combines Yolov3,Faster R-CNN,SSD target detection algorithm to implement pattern recognition and positioning.Comparative experimental results show that the accuracy of the improved VGGNET’s identification of clothing styles and colors reached 96.49%;YOLOV3 in the target detection algorithm of clothing pattern recognition and positioning reached 86.66%.The effect is the best,its MAP is 96.14%,the animal pattern MAP is83.69%,and the text pattern map is 79.80%.In terms of clothing patterns,this article uses two ways to study clothing patterns to obtain generation.The first generation formula,this article proposes a clothing pattern automatic generating method based on the improved deep convolutional format(DCGAN)model.In view of the different requirements of local details and overall aesthetics,the clothing pattern is mainly from two aspects.Improve DCGAN: Use the Prelu activation function to replace the RELU activation function in the generated network to obtain the local details and laws of rich features to increase texture patterns;Essence The introduction of objective and subjective ways to evaluate the generated patterns.The results show that compared with the traditional DCGAN,the two improved models proposed have better comprehensive performance,and the overall FID is reduced.Textile and clothing professionals also have a good evaluation of the quality,diversity,attractiveness and inspiration of the pattern generated by the improved model,which has high practical significance and application value.The second transformation formula,this article proposes a smart design and generation method combined with semantic segmentation and style migration technology.The previous study found that the patterns generated by the formaty method of the figure pattern are poorly effective.Then introduce the similar cat patterns as research objects,and select three styles: ink and ink style,oil painting style and pixel style as the style standard.In view of the characteristics of different elements of clothing patterns,build Deep Lab V3+semantic segmentation models to divide the target patterns and backgrounds,and set comparative experiments to set up different feature extraction networks.The style migration part is to build a fast-style migration model,and set the comparative experiments of three sets of different content loss and style loss parameters.The experimental results show that compared to the semantic segmentation model of the Xception main network,the semantic segmentation model of the Mobile Net V2 main network shows a better segmentation result.At the circumstances of both response,the MIO and MPA values have reached 94.12%and 96.98,respectively.%.In the style migration experiment,the results of comprehensive personal analysis,objective quantitative evaluation and subjective qualitative evaluation results,the best parameters of ink and ink style patterns are α = 1E5,β = 5E10;= 1E5,β = 1E10.Finally,in the method of instance verification,input the cat pattern into the fast-style migration model with the mask image obtained by the semantic segmentation model to generate a mixed-style clothing pattern. |
PDF Full Text Request