Research On Recognition And Classification Of Woven Fabric Pattern Using Deep Learning And Feature Fusion

The woven fabric pattern is the most important factor for woven fabrics that plays a significant role in the design,quality control,and production.The fabric properties,such as bending,stiffness,and firmness,as well as the appearance of woven fabrics,are influenced by the weave pattern.Woven fabric pattern recognition and classification is a critical task for massive production and quality control.Traditionally,woven fabrics are still recognized manually by human eyes or with a magnifying glass in the textile industry.This inspection is performed by an expert who requires extensive experience.However,it is labor-intensive,inefficient,and time-consuming.It also leads to subjective human factors,such as mental and physical stress,dizziness,tiredness,etc.,which ultimately affect the recognition results.Moreover,traditional algorithms need a lot of manual adjustments,which are time-consuming,complicated,and error-prone processes.Woven fabric pattern recognition still faces a lot of challenges and difficulties in improving the accuracy and robustness.The unique characteristic of woven fabric,that is,the periodic structure of woven fabric,physical properties of fabric,such as variation in fabric thickness,yarn diameter,and rotational variation of fabrics,have a direct influence on the recognition accuracy.Woven fabric color information is also very crucial because woven fabrics are available in solid and multi-colors.The arrangement of these color yarns and weave pattern increases its importance for woven fabric pattern recognition.Furthermore,color information plays a critical role in an industrial environment where illumination conditions are poor,and fabric analysis is conducted in random illumination.Therefore,to address these challenges,this thesis aims to consider all these characteristics to develop an efficient and automated method for the recognition and classification of woven fabric pattern by using deep convolutional neural networks(CNN)combined with wavelet and fabric color features.Firstly,we created a comprehensive woven fabric image dataset.The woven fabric pattern recognition and classification task needs a strong dataset of fabric images having distinctive samples of textured patterns varying in terms of weave style,fiber content,and different colors.There are no such sources available where we can find the dataset publicly.The dataset was created using a digital camera surrounded by light sources to control the light distribution.The dataset was generated in two illumination conditions: uniform and non-uniform(random)illumination.The images were captured from different locations on the fabric samples.The fabric samples were collected from several textile industries and warehouses and had a high variation in their weave patterns,fabric color,yarn diameter,fabric thickness,and fabric orientation.During image acquisition,fabrics contain non-uniform illumination,skewness(rotational variation)of the fabric,low-contrast images,and noise due to the surroundings and hardware,which is inevitable but affects the recognition accuracy.In view of the above situations,several preprocessing steps are carried out to create a better dataset.In a real-world textile industrial environment,a problem of specular reflection occurs,resulting in the white light spots on the fabric images,so a technique called “specular reflection removal” was performed to remove this non-uniformity of the illumination.The skewness of the fabric images caused during image acquisition was corrected and aligned using the Hough transform.Contrast limited adaptive histogram equalization(CLAHE)was employed to reduce the uneven distribution of gray level pixels caused by the light source,which helps in improving the contrast of the images.Some random noises are added during digital image acquisition due to environmental factors.It is necessary to remove the noise in advance,so a total variation denoising filter was implemented to clean the noise while preserving the important information about the edges.Due to a limited number of sample images,several data augmentation techniques were also performed to increase the size of the dataset.Secondly,we integrated the unique characteristics of woven fabric periodic texture structure by developing a model combining CNN and wavelet information for woven fabric pattern recognition and classification.Initially,we propose a CNN-based model that is used for extracting the CNN-based features,and then the classification of woven fabrics is performed.Then,taking advantage of the woven fabric periodic nature that is formed by the interlacing of warp and weft yarns,we implemented the discrete wavelet transform to extract the wavelet textural features.Next,we combine the proposed CNN-based model with the discrete wavelet transform method to improve the overall performance of our model.The fusion of the CNN-based model with the discrete wavelet transform incorporated more detailed features from the fabric images to recognize the weave pattern accurately.We tested our proposed model using cotton woven fabrics because they are widely popular and versatile.To validate the performance of our model,several parameters and evaluation metrics were evaluated.The experimental results show that the fusion of CNN and wavelet approaches outperformed existing studies significantly.Thirdly,we incorporated the woven fabric color information by proposing a fusion model that employs a transfer learning(TL)-based deep CNN(modified Res Net-50)model and fabric color features for the woven fabric pattern recognition and classification.We used a residual network based deep CNN model using Res Net-50 as a base model,with newly added layers at the top of the architecture.Transfer learning is used to transfer the weights and parameters from the pretrained model to our modified deep CNN model.This model uses the characteristics of residual network for the extraction of deep CNN features that enable us to deal with the fabrics’ having variation in their physical properties.We used the color moments to extract fabric color features.Then,we implemented the fusion of transfer learning-based deep CNN features and fabric color features.The developed fusion model was tested over the woven fabric dataset,which contains many kinds of woven fabrics.The performance of the proposed method was evaluated using various performance metrics.A comparative analysis was carried out against other baseline approaches and existing works.The experimental results show that the proposed fusion method achieved higher classification accuracy and performed better than the existing studies.Finally,considering the real-world industrial situation,a comprehensive fusion model that incorporates wavelet information,fabric color information,and deep CNN features was proposed.The woven fabric pattern recognition becomes challenging in an extreme industrial environment where the illumination conditions are poor and fabric analysis is conducted in random(nonuniform)illumination.In such industrial environments,the unique characteristics of woven fabric periodic structure,fabric color information,and physical properties of woven fabrics must be considered altogether.Therefore,we deal with the three color spaces that are RGB,HSV,and CIE LAB,for the fabric color features extraction.The feature parameters computed for these three color spaces are mean,variance,standard deviation,skewness,and kurtosis.Discrete wavelet transform is used to extract the wavelet textural features.The TL-based deep CNN(modified Res Net-50)model developed and discussed in the previous paragraph is used to extract the deep CNN features.Once these aforementioned features are acquired,they are concatenated together to contain all the features of woven fabric images.Furthermore,significant features were selected by the genetic algorithm for the classification of woven fabrics.We performed several experiments to validate our proposed fusion model,and the classification results demonstrated that the significant features selected by the genetic algorithm have performed remarkably well.The proposed model was also compared with existing approaches.The experimental results show that the proposed method is efficient and outperforms existing studies,and capable of performing with high accuracy in harsh industrial environments.