A Study On Fabric Drape Based On Three-Dimnesional Model And Deep Learning

Fabric drape refers to the ability of a fabric to form a three-dimensional(3D)configuration under its own gravity.Studies on fabric drape is beneficia to fabric development and garment design.However,there are several topics on fabric drape have not been properly investigated,including the extracting of fabric drape indicators based on 3D drape models,the resampling of 3D drape models,the reconstruction of 3D drape models when three-dimensional scanning is not feasible,the matching of 3D drape models,and the 2-way mapping of fabric drape and fabric mechanic properties.Therefore,3D fabric drape model was used as the raw information in this thesis to address these topics with the help of machine learning and deep learning.(1)The 3D point cloud of draped fabric was captured via a self-built 3D scanning device followed by surface triangulation.a new drape indicator(Drape Angle,DA),which can be used to evaluate the drape ability of the draped fabric,was proposed based on the statistics of triangle normal of the 3D drape model.The comparison between DA and drape coefficient(DC)was conducted.The results show that when there is self-overlapping in the drape,DA is more objective and persuasive in charactering fabric drapeability than that of DC.Besides,the CV(coefficient of variation)of DAs is less than that of DCs for the same specimen.(2)In order to standardize the 3D drape model collected by the 3D scanning device,a new method based on the local linear embedding and shared triangle-weights was proposed to resample 3D drape models.Firstly,the 3D drape model was mapped into a two-dimensional(2D)triangular mesh in the shape of a circle.Secondly,the mapping result was resampled by a 2D point cloud uniformly distributed.Thirdly,the triangle weights of the 2D triangular mesh resampled by the 2D point cloud was used to infer the resampled points on the 3D drape model.The results show that in the process of mapping the 3D drape model into 2D triangular mesh,there is no obvious tensile deformation,which makes the resampled points from the 3D drape model evenly distributed.After surface resampling,all drape models possess the same triangular topology,which makes the comparison among different drape models more efficient.(3)A neural network formed by a circular template triangular mesh,a convolutional neural network with residual modules,and a graph convolutional neural network as the deformation module was designed.The convolutional neural network with residual modules was used to extract the embedded feature of a single image.Under the constraint of the embedded feature,the deformation module was used to change the 3D configuration of the circular template triangular mesh.With this neural network,a 3D drape model could be reconstructed with only a single image as the input.All reconstructed 3D drape models have the same triangular topology.The mean distances between the reconstructed vertices and the true vertices is 2.3031mm.The mean absolute relative error of drape angle is 2.7099%,which means the proposed method could be used under the situation where 3D scanning is not feasible.(4)For the matching of 3D drape models,three different methods were proposed,i.e.,matching by the principal component of fabric drape indicators(Ipca),matching by the principal components of curvature of fabric drape model(Cpca),and matching by the combination of Ipca and Cpca,i.e.,ICpca.It is proved that for the same dataset,the recall rate of fabric drape model was ranked as Ipca<Cpca<ICpca.When Ipca and Cpca were combined as the principle ICpca=[?1·Ipca(?)(1-?1)·Cpca]((?)refers to concatenation),and ?1 is set between 0.855 and 0.895,the recall rate is relatively high.(5)To explore the 2-way mapping of fabric drape and fabric mechanic properties,thereby to probe a possible method to predict fabric mechanic properties with 3D drape model,BP neural networks and deep learning were used to tackle two problems,i.e.,the regression of fabric mechanic properties and the classification of fabric softness.In the mapping of fabric mechanic properties to drape configuration,the prediction of fabric drape coefficient(DC)as well as drape angle(DA)were conducted by feeding a BP neural network the combined features consist of fabric weight(Gk),warp bending rigidity(Bwarp),weft bending rigidity(Bwef t),warp shearing stiffness(Swarp),and weft shearing stiffness(Seft).In the mapping of fabric drape configuration to fabric mechanic properties,the prediction of Bwarp,Bweft,Swarp,Sweft were conducted by feeding another BP neural network the combination of ICpca and Gk.The results show that the correlation coefficient between the predicted Bwarp,Bwef t and their true values are higher than that of predicted Swarp,Sweft.Apart from that,the mapping of drape configuration to fabric mechanic properties could be treated as a task of classification.Fabric samples were divided into different classes in terms of their?b/g(the ratio of Bwarp+Bweft to Gk)guided by K-Means clustering,i.e.,all fabrics were labelled according to their ?b/g.Five deep learning models were trained with transfer learning to classify the grey images of draped fabric.The accuracy generated by the best deep learning models of binary classification is equivalent to 94.43%to 94.58%of the accuracy tested by humans.And the performance of the best deep learning models in the task of classifying the softness into three classes is equivalent to 96.8%?97.8%of the accuracy tested by humans.