Deep Learning-Based Large-Scale Color Fabric Pattern Decoding

With the popularization of intelligent manufacturing,computer techniques have been broadly employed in production to save labor and time costs.Intelligent jacquard looms weave specific coded patterns by loading binary matrices that simulate fabric structure.Using computer vision to generate binary matrices from fabric patterns can help designers and producers save much time and energy in reproducing and protecting fabrics from being dissected.In generating binary matrices for complex jacquard fabric patterns,an algorithm has proposed an intermediate representation for labeling the location and type of crossing points.Based on the input textile pattern and the UNet network model,the algorithm predicted the intermediate representation and generated the binary matrix using post-processing techniques.However,noise interference affects the predicted intermediate representation of the algorithm,and its prediction accuracy must be improved.Moreover,because of the constraints with the dataset,the algorithm is only capable of predicting specific small-sized grayscale fabric patterns,and cannot adapt to arbitrary-sized color fabric patterns.To tackle these challenges,this thesis focuses on improving network model and optimizing algorithmic processes,the details are as follows:(1)To improve the accuracy of automatic detection of the crossing points and to address the issue of large noise in the results,this thesis proposes to use the Res UNet network structure as the backbone network,and introduce the Convolutional Block Attention Module(CBAM)attention mechanism to reduce the noise generated by simple concatenation of low-level features and high-level features,more accurately suppress unnecessary features and highlight meaningful features.Meanwhile,the atrous Spatial Pyramid Pooling(ASPP)is introduced to replace the bridging part between the encoder and decoder,and the multi-scale atrous convolution feature extraction module is modified to a dense connection to fully integrate the contextual information of the fabric at different scales,making it more suitable for jacquard fabric pattern decoding.Finally,a hybrid loss function that combines Mean Square Error(MSE)and Structural Similarity Index Measure(SSIM)is introduced to increase the prediction accuracy,which can consider both the pixel-by-pixel similarity and the similarity of the local area of the predicted intermediate representation.(2)To address the input restrictions of the algorithm for automatic detection of crossing points,this thesis first adopts the image stitching algorithm based on feature point matching to obtain high-resolution color fabric pattern of arbitrary size.Then,to tackle the issue of poor prediction accuracy of the crossing points caused by the similar brightness of weft and warp colors,this thesis proposes a high-resolution color fabric pattern grayscale conversion method based on principal component analysis and utilizes the overlap-tile strategy to crop it into sub-patterns with the same size as the network model input for crossing points prediction.Finally,this thesis infers the receptive field size of the feature maps in the neural network structure,calculates the confidence level of each pixel in the crossing points prediction map,and employs a weighted fusion method to stitch together all the crossing points prediction maps to obtain the crossing points prediction map of high-resolution color fabric patterns of arbitrary size.This thesis addresses the drawback of the previous algorithm that cannot predict large-scale color jacquard fabric patterns.