Fabric defect detection based on feature fusion of a convolutional neural network and optimized extreme learning machine

Aiming to accurately detect various defects in the fabric production process, we propose a fabric defect detection algorithm based on the feature fusion of a convolutional neural network (CNN) and optimized extreme learning machine (ELM). Firstly, we use transfer learning to transfer the parameters of the first 13 convolutional layers and first two fully connected layers of a VGG16 network model as pre-trained by ImageNet to the initial model and fine-tune the parameters. Subsequently, the fine-tuned model is used as a feature extractor to extract features of RGB images and their corresponding L-component images. A principal component analysis is used to reduce the dimensionality of the features and fuse the reduced features. The moth flame optimization (MFO) algorithm is used to initialize the optimization variables of a parallel chaotic search (PCS) algorithm, and the PCS algorithm (as optimized by the MFO algorithm) is used to optimize the input weight and bias of the ELM (i.e., the PCS-MFO-ELM (PMELM)). Finally, the PMELM is used to replace the softmax classifier of the CNN to classify and detect fabric defect features. The experimental results show that on the amplified TILDA dataset, the precision, recall, F1-score, and accuracy rates of this algorithm for fabric holes, stains, warp breaks, dragging, and folds in fabric can reach 98.57%, 98.52%, 98.52%, and 98.50%, respectively, that is, higher than those of other algorithms. Through a validity experiment, this method is shown to be suitable for defect detection for unpatterned fabrics, regular patterned fabrics, and irregularly patterned fabrics.