| Clothing patterns can be expressed in the form of raster graphics as well as vector graphics.In production practice,it is inevitable to automatically convert the former to the latter.In other words,as one of the common key technologies in the digitization process of clothing patterns,it is imperative to study the vectorization technology of clothing-pattern raster.In view of this,this paper uses deep learning to realize the input clothing-pattern raster,automatically output the corresponding clothing pattern key point coordinates,and then draw the digital clothing pattern.The specific research work is as follows.(1).Build a pattern dataset driven by clothing parameters.In this paper,taking women’s tops and women’s pants as the research object,and in order to automatically generate a variety of version data sets,the relevant clothing parameters of five clothing versions of the top front piece,top back piece,trouser front piece,trouser back piece and sleeve are extracted respectively,and different clothing parameter size ranges are set for different styles,and random values are taken from the design range as the final clothing parameters of a certain clothing version.Secondly,using the Bezier parameter equation,De Casteljau and other algorithms,the key point coordinates of the version are calculated based on the above clothing parameters,and Python libraries such as Cairo and Cairo SVG are used to draw the clothing pattern grid diagram according to the coordinates of the key points.The final dataset contains the raster and the key point coordinates parameter for the corresponding layout.(2).The neural network based on Dense Net121 realizes the automatic generation of digital clothing patternsThis chapter first completes the normalization and data augmentation preprocessing of the built dataset,takes the clothing raster pattern in the dataset as the input of the neural network,uses the Dense Net121 neural network to extract the features of the raster graph with fewer parameters and higher efficiency through multi-layer convolution,pooling operation of dense block and transition layer,and uses the fully connected layer to return the final feature to the target parameter.That is,the key point coordinate parameters of the clothing version,the weighted MSE function is designed to update the model weight and bias,and finally the predicted Bezier parameter coordinates are further automatically drawn for the clothing version.Among them,5 different kinds of clothing versions are regarded as different training tasks,and the same Dense Net121 neural network architecture is used to modify the number of neurons in the final fully connected layer(the number of key coordinate parameters corresponding to the version)to achieve the task of generating different types of clothing patterns.(3).Automatically generate parametric clothing patterns based on vectorized raster graphsFrom the research of existing vectorized raster maps,by testing the clothing pattern drawings in three different scenarios(non-rotating clothing raster pattern,rotating clothing raster pattern,scanned clothing raster pattern),the neural network model suitable for automatically generating digital clothing pattern diagram was selected,and finally the CNN-RNN-based neural network framework was selected,the raster graph in the self-constructed dataset was used as input,and the Bezier parameter curve was used as the key point coordinates of the parametric clothing pattern.After the prediction results are further automatically drawn into the clothing pattern diagram,for the three defects that may appear,one is that there are redundant and repeated stitches,the second is that there are prediction errors in the curve,and the third is that there are no C~1 continuous defects between the curves.The identification and post-processing are carried out to optimize digital clothing pattern.Finally,the model prediction and optimization results are evaluated using Hausdorff distance and F-score. |
