DEFECTCNN: Improved Discriminative Convolution Neural Network Towards Instantaneous Automatic Detection and Classification of Complex Defect in Fabrics

Abstract

                  Due to enormous growth of textile industries has increased demand for the automatic fabric defect detection and classification system to the fabric material as it plays a crucial role in maintaining the quality of the services. Machine learning model has employed as automatic defect detection system to identify the material quality. Despite of several advantageous of the machine learning model, those models faces several challenges on handling the complex and uncertainty of varied texture and structural patterns. Further it is complex to process the boundaries and features with high degree of intra class variation and low degree of interclass variations. On leveraging and exploiting the deep learning architecture, the over lapping and varied texture patterns can be efficiently discriminated on defects. In this paper, a new deep learning architecture entitled as discriminative convolution neural model is proposed to detect and classify the defects in the fabric materials into various defect classes. Initially fabric image preprocessed on basis of the noise filtering through wiener filter and image enhancement through CLAHE technique. Enhanced image is segmented using image thresholding technique to segment it into the various regions on basis of pixel information’s by grouping the neighbouring similar pixels intensity or textures to represent a mask. Segmented image regions are projected to the convolution neural network. Convolution layer of network is to extract the features from its composition containing kernels with different weights. It computes the high level features for different pixels based on surrounding and neighbouring pixel values on striding to produce the feature map containing gradient and edge of the images.  ReLU activation function is applied to reduce the non linearity among the features in the feature map. Pooling layer of the model down-sample the convolved features to produce the activation map. Activation map is obtained using max pooling as it returns maximum value from the segment of the image processed using kernels. Activation map is transformed into tabular structure to perform the classification easily. In addition drop out layer is incorporated in the model to eliminate the overfitting issue during classification on reducing the correlation among the neurons. Fully connected layers of the model is used to learn the flattened features with weights and bias to classify the flatten features using softmax layer on basis of defect classes such as Hole , Color Spot, Thread Error  and foreign body.  Experimental analysis of the proposed architecture is carried out on TILDA dataset using cross fold validation to analyse the representation ability to discriminate the features with large variance between the different classes. From the results, it is confirming that proposed architecture exhibiting higher performance in classification accuracy of 98.43% in classifying the fabric defect on compared with conventional approaches