Research On Fabric Appearance Defect Detection Based On Convolution Neural Network

With the development and progress of society,textile technology is improving day by day,and textile products can be seen everywhere in daily life and industrial scenes.While textiles are widely used,they also have higher requirements for their quality.In the fabric production process,product quality inspection is a very important step.The timely detection and treatment of defects such as surface breaks,wear,scratches,holes,and oil stains on the fabric surface can create value and reduce losses for enterprise production.At the same time,fabric images have complex backgrounds,subtle defects and difficult to detect.Moreover,manual inspection and traditional methods are inefficient,complicated in execution,and ineffective.There is an urgent need to develop more intelligent and efficient fabric appearance defect detection methods.This paper aims at overcoming the disadvantage that the effective classification and segmentation of fabric defect images cannot be realized.Combined with advanced computer vision algorithms,a CNN-LSTM-based defect classification method and a CNN-based multi-branch parallel defect segmentation method are proposed.The main details and achievement in the dissertation are as follows:(1)Using the labeling software Labelme to label the image to solve the problem of the lack of correct labels in the fabric image.Using data enhancement to solve the problem of fewer fabric images.(2)When the conventional convolutional neural network is used for classification tasks,the classification effect is not ideal on some data sets where the difference between defective images and normal images is small.This paper proposes a defect classification method based on CNN-LSTM,which effectively combines CNN and LSTM,so that the network can simultaneously extract spatial features and potential connections between images to obtain more information.The steps of the method of CNN-LSTM which based on extracting spatio-temporal characteristics are as follows: first use the classic CNN network to extract the spatial features of the fabric image;then put the extracted spatial features into the LSTM network to extract the latent relationship between the images,and complete the space-time features Extraction;Finally,the features obtained by the above two steps are sent to the classifier to complete the classification and recognition.In addition,considering that the structure of CNN-LSTM will make the network very deep,a smoother Mish activation function is used.After experimental verification,this combination of CNN-LSTM can significantly improve the classification accuracy: the accuracy rate on the DAGM data set is99.81%,and the accuracy rate on the TILDA data set is 98.81%;Compared to those of a single convolutional neural Network like Inception,Res Net and Mobile Net,the method in the paper has better classification effect.(3)The conventional segmentation network only uses superposition for convolutional layers of different resolutions to operate in series.This paper proposes a multi-branch parallel network model.The network model is composed of three modules,which is backbone module,the attention module and the feature fusion module.The backbone module is formed by paralleling six branches whose resolution decreased layer by layer.The feature map of the main branch maintains the original resolution,and the feature maps of the other five branches will be merged into the main branch output through feature fusion to complete the multi-scale fusion of different branches;The attention module uses the channel attention mechanism to recalibrate the original features according to the feature importance;The feature fusion module is designed to keep the balance between spatial information and receptive fields,and integrate spatial information and semantic information effectively.The multi-branch parallel network model greatly improves the fineness of segmentation,which is closer to the real label.After experiments,it has obtained 99.48% pixel accuracy and89.09% IOU score on the DAGM data set,99.23% pixel accuracy and 76.89% IOU score on the TILDA data set.Achieved better segmentation results than other network like Deep Lab V3+,UNet,PSP,FCN and Refine Net.