Research On Automatic Evaluation For Fabric Pilling Based On Depth Information Of Fuzz

Traditionally,the evaluation of pilling performance depends on manual visual judgment,bycomparing a pilled fabric with stand sample photographs.The visual measurement method is very inefficient and easy to be affected by subjective factors,so the result sometimes is unreliable.As the development of automatic control and computer vision,researches on objective and accurate pilling assessment have arisen.The thesis is aimed at studying pilling-evaluating algorithms based on image processing and eventually developing an automatic pilling-grading system.Present automatic pilling detection are mostly based on two-dimensional gray-scale images or three-dimensional images.Pills can be extracted according to different gray values between fuzz cluster and fabric background.However,since the gray-scale threshold method can be easily disturbed by irregular patterns,it is only suitable for plain colored fabrics with periodic texture.In addition,the two-dimensional images can't show the height data of pills,which is a key indicator for segmenting and describing pills.The three-dimensional images can be obtained by laser scanner,binocular stereo equipment or other devices such as the self-made device by Xia Chen for capturing lighting projected images.However,the number of studies on three-dimensional images are much lower that two-dimensional images,due to the expensive and hardly-operated devices.The magnification level of most image capture devices are not enough to capture fuzzes on fabric surface,therefore most pilling assessment studies are based on pills,ignoring the fact that fuzzes are critical factors for performance evaluation.Based on above analysis,the thesis was focused on the following two difficult points: to carry out the study of computer pilling evaluation based on fuzzes,which requires a suitable capture device with enough magnification for capturing fuzzes;and to obtain the height information of fuzzes,pills and fabric surface.Therefore,the technique of Depth From Focus(DFF)was adopted in this thesis,allowing the system to obtain the depth values of pills,fuzzes and fabric surface by focusing under microscope lens.Further studies on thesegmentation and characteristic parameters extraction were based on the depth information of fuzzes and fabric surface.The basic idea of this thesis was firstly reconstructing depth information of fabric surface and fuzzes from a sequential images captured under the microscopy at different focal positions,and then segmenting fuzzes and fabric surface by using predicted fabric surface base-level plane,thirdly extracting several features including coverage area ratio,coverage volume ratio,maximum height,concentrated height and roughness,then by using these characteristic parameters and SVM classifier,evaluating the pilling grades of fabrics under different friction time(number of friction turns),finally realizing the development of the automatic fabric pilling evaluation system.The objective evaluation results were compared with manual visual results to verify the accuracy of the system.The research included the building of automatic detection hardware system and the study of software algorithms.The major research contents included: the research on the method of calculating depth values of pills and fabric surface from the sequential layers of images;the research on the method of constructing the fabric surface base-level plane to extract fuzzes;the research on automatic evaluating the pilling grades by using the combined characteristic parameters and SVM classifiers.Furthermore,the research was carried out on dynamic description of fabric pilling performance by rubbing fabrics at different turns of friction.Several characteristic parameters,combined with the fabric pilling grading were used to represent the pilling performance.Curves of the characteristic parameters and pilling grades variable against friction time(number of friction turns)were used as the combined characteristic parameters for dynamic variation of pilling performance against friction time(number of friction turns).The main work and contributions of this thesis are as the following:(1)The design of the hardware system for automatic evaluation of fabric pilling and scheme for image acquisitionThe design of the hardware system consisted of several parts including illumination and data transmission between computer and the microscopic system.The back lights provided by halogen light of the microscope failed to illuminate thick fabrics.Therefore,a LED annular light used as forward lights was fixed around the microscopic lens.The computer could control the movement of objective stage through serial port.The images were input into computer by USB2.0 cable.By doing so,the system were able to realize automation of platform and acquisition of images.The design of scheme for image acquisition mainly include the acquisition of sequential layers of images and the movement of platform during multiple views scanning.By measuring the distances of each steps for platform along three directions(x,y,z)and field of depth,acquisition parameters including the depth distance between layers,the number of images and step lengthbetween views could be set.To eliminate the return stroke error caused by the rack and pinion drive during multiple views scanning,a special route of platform movement was designed.(3)The reconstruction of depth information based on sequential layersThe method of reconstructing depth values based on pixel sharpness was put forward in this thesis.Since the captured sequential layers could cover all the depth positions of fuzz and fabric surface,for any(x,y)position,one could find a corresponding focused point among sequential layers.Therefore,to calculated the depth value for one(x,y)position,all the points at plane coordinates(x,y)and different focal positions were scanned to find the one with maximum sharpness.The number of layer where the best-focused point was found could be recorded as the depth value.The clarity-evaluation index was very critical in this algorithm.Traditional sharpness evaluation functions were discussed firstly.By analysing the depth similarity between one pixel and its surroundings,a clarity-evaluation method based on variance of gradient within local region was put forward.Method of adaptive selection of local region was presented in the thesis.Experimental depth image showed complete form of the fuzz without rupture and breakage,and the variation of obtained depth values were consistent with the real height variation of fuzzes.(3)The segmentation method of fuzzes and fabric surface based on depth valuesThe segmentation algorithm was proposed based on the fabric base-level plane.The base-level plane was predicted by several basic points locating at fabric surface.Fuzzes above the predicted plane could be extracted By analysing the spacial distribution of fuzzes and fabric surface,the basic point were selected by using Meanshift technique and spacial vectors.After Meanshift segmenting,several patches were formed.Then,patches locating on fabric surface were extracted based on the threshold of depth values and spacial vectors.Finally,the predicted base-level plane was established by using these surface patches as basic points.The experiment result indicated that the predicted base-level plane could estimate the depth values of fabric surface well.(4)The pilling evaluation method by using combined features and modified support vector machine(SVM)classifierThe automatic pilling rating algorithm based on combined features and support vector machine(SVM)was discussed int the thesis.Five parameters including coverage area ratio,coverage volume ratio,maximum height,concentrated height and roughness were extracted as the input of SVM to output the pilling grades.216 testing areas on 72 colored woven and knitted fabrics with different colors and patterns were selected to conduct the experiment.Considering the limited view of microscopic,the number of views in detection area were discussed.The SVM classifier was modified by using grid optimization method.The accuracy rate of the model was up to 89.02% by the cross validation of 219 groups of combined features.(5)The dynamic description of pilling performance variable against friction time(friction turns)Present fabric pilling evaluation method is simulating pills under a certain turns of friction,and then giving a pilling rate.However,fabric pilling is a dynamic process,traditional evaluation approach can't fully describe the characteristics of fabric pilling.In order to observe the pilling condition at each friction stage,several observation points of friction turns were set on the Martindale pilling tester.In addition,three features including the coverage area ratio,the coverage volume ratio and the roughness were calculated and the curves of features vary with the friction turns were observed.Through automatic grading of fabrics,the curve of pilling grade over friction turns was established.The objective results were verified to consist with manual vision assessment.Finally,the above four curves were used to describe the dynamic pilling performance variable against friction time(friction turns).In the thesis,an automatic pilling grading system was developed based on the depth information of fuzzes and fabric surface,and the dynamic description of pilling performance vary with the friction time was studied,laying an important theoretical fundation for establishing an overall,objective pilling evaluation system.