Machine Evaluation Of Fabric Smoothness Appearance Based On Image Analysis

The fabric smoothness appearance after laundering is an important indicator of the wrinkle resistance of the fabric,and the wrinkle resistance is one of the main determinants of the shape retention of the fabric,which plays an important role in textile production,trading,and use.At present,for the fabric smoothness appearance evaluation,manual subjective evaluation method is still the dominant position in the textile industry.However,the subjective method depending on human vision has many shortcomings,such as low accuracy,poor stability,weak objectivity,high cost,and irreproducibility.For this reason,objective methods that can avoid the aforementioned shortcomings have been one of the hot research issues.The current research mainly uses computer vision technology to achieve the objective evaluation of the fabric smoothness appearance.By the form of the fabric appearance data,the methods of obtaining the fabric surface data can be divided into two-dimensional imaging method(2D method)and three-dimensional imaging method(3D method).Among them,the 2d method is susceptible to the interference of fabric color pattern and image acquisition environment,while the 3d method has the disadvantages of low efficiency and high calibration complexity.After a comprehensive analysis of the advantages and disadvantages of them,this thesis proposes to improve the evaluation performance of the 2D method through the optimization of the image acquisition system and the optimization of the computer vision algorithm.The main work is as follows:(1)In order to improve the environmental stability of the 2D method,a fabric 2D image acquisition platform is first constructed with the image acquisition parameter optimized.The parameters including altitudinal angle,position angle,and brightness of the light source illumination of the platform can be controlled.Under different environmental parameters,fabric sample images are collected to construct a data set containing 385 fabric samples with different smoothness levels.A series of commonly used image feature sets are extracted and whose information gains are analyzed,which determines the optimization priority of each parameter.The support vector machine(SVM)model is trained with each set of features,and the test performance of the model is used to evaluate the expressive ability of different environments.According to the determined priority order,the optimal image acquisition environment is obtained.(2)This thesis studies the possible schemes to improve the adaptability of the 2D method to the illumination environment and the color pattern of the fabric.The photometric model of the fabric surface is established,and the composition mechanism of the general illumination,color pattern,wrinkle pattern in the fabric image is obtained.Based on this model,polynomial fitting method is used to remove the general illumination from the non-color pattern fabric image.Furthermore,based on the photometric theory,a fabric color texture shading model is constructed to generate pairs of non-color and color pattern fabric image pairs to train a supervised image translation model to remove the color pattern.The experimental results show that the supervised image translation model performs significantly better than the unsupervised model.In addition,in the real color pattern fabric images,the model can maintain the accuracy of fabric smoothness evaluation under the level 0.5 error to more than 75%.(3)Based on the related research of the human visual system(HVS),a multi-scale visual masking model is proposed to extract the low-level visual features of the fabric image.This model simulates the bottom-up visual perception of human vision,and can effectively describe the characteristics of HVS such as contrast sensitivity,multi-channel characteristics,visual masking,etc.,which are highly related to the visual perception of the fabric smoothness appearance.Within the collected image database of fabric images,the SVM multi-classification model is trained with the above image features,and it shows better test accuracies than existing methods,which verified the effectiveness of the proposed low-level features.(4)According to the characteristics of fabric smoothness appearance evaluation task with a small sample size and low abstraction,a compact convolutional neural network(cCNN)is designed to extract high-level visual features of fabric images.The model simulates the bottomup visual perception mechanism of HVS.In comparison experiments,the cCNN model achieves better overall performance than existing methods,which verifies the effectiveness of high-level features in describing the smoothness appearance of fabrics.Considering the ordinal classification attribute of the problem,a label smoothing objective function is proposed.The ordinal label prior of the problem is introduced,which effectively improves the performance of cCNN.(5)On the basis of the visual features in different levels,a multi-level compact convolutional neural network(McCNN)is proposed,and the results of this research are synthesized to construct a set of objective evaluation system for fabric smoothness appearance.The McCNN refers to the complementary experiment of high-order features and low-order features,and simulates the hierarchical information perception structure of HVS.In addition,considering the possible sample label errors caused by subjective evaluating,a mislabeled sample filtering strategy is proposed to mitigate the negative impact of mislabeled samples on model training.In the performance verification experiment,the McCNN model achieves accuracies superior to the existing methods,which are 86.06%,96.28%,and 100% under the errors of 0,0.5,and 1 level respectively.It demonstrates the desirable industrial application value of the proposed McCNN.Based on the above research work,this thesis proposes a two-dimensional image-based fabric appearance evaluation system.The system has good image acquisition environment stability and sample color pattern adaptability.It adopts image features and evaluation models that conform to the HVS perception mechanism,and presents the industrial application prospect of the objective evaluation method of fabric smoothness appearance.