Detection of Stain Defects in Textile Industry using State-of-Art Transfer Learning Models

Manual detection of fabric defects in textile industry is a cumbersome & fatigue job and quite often prone to error, henceforth there is a need for programmed defect detection as part of quality control process. For any production unit, the Quality control is of paramount prominence. If defects are not detected, it results in financial loss and adverse impact on reputation in market. Image processing is key methodology for detecting defects in fabric; presently deep learning is immensely applied for effective image analysis. Fabric stains are one of the common types of fabric defect. This work, targets to automate the task of detection of three types of stains: ink, dust and oil in fabric using Convolutional Neural Network (CNN). One of the major issues with CNN is, it requires large input dataset for better performance. To tackle this issue as well to reduce training time and improve classification accuracy, four transfer learning methods: VGG16, Xception, MobileNet and InceptionV3. Results are compared with basic model of CNN. Among the various pretrained model, the MobileNet resulted in highest accuracy of 97.14% and 97.64% with training and validation dataset respectively with just five epochs; VGG16 resulted in lowest training and validation accuracy of 88.81% and 90.17% respectively with 10 epochs.