Research On Microscopic Vision Feature Representation And Identification Algorithm For Cashmere And Wool

The morphological characteristics of cashmere and wool fibers,as well as their physical and chemical properties,are similar,so the identification of these two fibers has always been a challenging proposition.Although many new methods were developed,the artificial detection method is still the most important practical detection method at present.However,the artificial detection method is time-consuming and laborious as it requires microscopists observe the morphology of the fiber under a microscope for a long time.This method is subjective and the detection accuracy is dependent on the experience of microscopists.For many years,researchers have been seeking to develop a fast and accurate method for automatically identifying cashmere/wool fibers and have made many useful attempts.Recently,computer vision and machine learning technologies have made great progress.In particular,deep learning has a very good application effect in image classification and object detection.In this paper,we mainly use computer vision technology to study the identification method of cashmere/wool fiber and the visual feature representation of fiber image.The main work and contributions of this thesis are as follows.(1)Base on the experience of manually identifying cashmere/wool fibers,various microscopes were used to acquire fiber images.By comparing and analyzing the factors such as the quality,efficiency and cost of obtaining fiber image,we have selected devices that are suitable for the rapid acquisition of fiber images and set the imaging standard.We took more than 50,000 microscope images and established a dataset of cashmere/wool fiber images.It provides a large sample support for further research on identification of cashmere and wool fiber based on visual morphology.(2)We proposed a method to identify fibers based on projection curve.First,the microscopic images of wool/cashmere fiber were transferred into projection curves.To extract the features of the projection curve,three different feature extraction methods,recurrence quantification analysis(RQA),direct geometrical description(DGD),and discrete wavelet transform(DWT)were employed to reveal the embedded numerical features.The extracted parameters were used to screen the supervised classification methods,including artificial neural network(ANN),kernel ridge regression(KRR),and support vector machine(SVM).After testing various machine learning methods,it was found that the best classification result could be obtained via the SVM trained by the data generated from RQA and the recognition rate reached 90.8%.We investigated the key parameters,threshold distance defined in RQA.It was found that the best classification result could be obtained when threshold distance was set to 5.(3)We studied the application of spatial pyramid matching and bag-of-words model in the identification of cashmere/wool fibers.In this paper,first,the fiber images were preprocessed and then the SIFT descriptor was used to extract features from fiber images.These features were employed to generate a codebook and every image can be described as a collection of codewords.Next,spatial pyramid matching was used to enhance the ability of codewords to express the spatial information of the image,and each image was transformed into a set of vectors.Finally,the vectors were fed into classifier for a supervised classification and the highest identification rate exceeded 93%.We investigated the two key parameters,codebook size and resolution level.By trade-off between discriminability and generalizability,resolution level and codebook size were set to 2 and 600,respectively.Comparing the recognition rate before and after fiber image preprocessing and it is found that removing background is beneficial to improve the performance of the model.To evaluate the stability of the method,we prepared fifteen groups of samples with different blend ratios.Identification accuracy of each group exceeded 90%.(4)We proposed a fiber identification model based on Local Binary Pattern(LBP).In this paper,the scale pattern of the fiber surface is considered as a texture feature,and LBP descriptor was used to extract features from fiber images.Each fiber image was converted to a vector,which is a histogram of LBPs extracted from fiber images.The vectors were input into SVM for a supervised classification.In the experiment,we compared several different LBP descriptors and found that Rotation Invariant Co-occurrence LBP(RICo LBP)and Rotation Invariant Co-occurrence among Adjacent LBPs(RICoALBP)can represent fiber features well.The former has a stronger description ability,while the latter has a lower feature dimension.The models were tested on scanning electron microscope images and optical microscope datasets respectively.The accuracy was about 96% and 90%,respectively.(5)Referring to the VGG-16 model,a Convolution Neural Network,Fiber-Net,was established to identify cashmere and wool fibers.Five different kinds of cashmere /wool fiber images in the data set were fed into the model,and the total recognition rate was 92.74%.Among them,Mongolian brown cashmere has the highest recognition rate of 99.7%,followed by the Chinese indigenous wool with a recognition rate of 96.7%.And the recognition rate of Chinese white cashmere was relatively low,reaching 86.4%.Recognition rates of Chinese grey cashmere and Mongolian grey cashmere were 90.8% and 89.8%,respectively.Experimental results show that our model can solve the problem of multi-classification fiber recognition.Compared with other methods proposed in this paper,Convolution Neural Network has the best recognition effect.(6)Taking VGG-16 model as an example,the application of transfer learning in fiber recognition is studied.The image features pre-trained on Image Net dataset were transferred to VGG-16 for the task of fiber identification.In the experiment,the parameters of the pre-training model were loaded layer by layer,and it was found that the effect of the previous feature transfer learning was better,while the high-level feature transfer significantly reduced the recognition rate of the model.The effects of transfer learning were evaluated on datasets of different sizes.Experiments have shown that transfer learning can improve the performance of the model when the dataset is small.