Fabric Defect Recognition Using Extremely Efficient Convolutional Neural Networks

The cost of advanced textiles is often affected by the fabric defects that represent a major problem to the garment industry,so the automatic inspection systems is are required by the textile manufacturers to maintain the fabric quality.The traditional machine vision and image processing methods are mostly used in existing inspection methods,while the traditional fabric defect detection methods based on machine vision have some limitations.Not suitable for detecting fabrics with rich textures and patterns.In this thesis,the latest research results in the field of artificial intelligence are combined with the needs of the industry,and the intelligent recognition method of fabric defects is deeply studied.The defect is robust and supports the online learning of a new lightweight convolutional neural network structure.The model greatly reduces the dependence on hardware computing ability and memory capacity while maintaining high recognition accuracy.It makes the deep neural network easier to run in the industrial field.At the same time,this study also provides a new way to solve the problem of intelligent identification of surface defects of other industrial products.The main research results of this paper are as follows:1)This thesis analyzes the deficiency of traditional feature extraction model,introduces in detail the structure configuration and training method of a classical convolution neural network model,and uses the convolution neural network to realize defect recognition of complex pattern fabric.To some extent,it improves the problem that the complex pattern fabric can not be detected effectively,and provides a theoretical basis for the follow-up research.2)Although the performance of convolution neural networks is very powerful,deep neural networks need to consume a lot of hardware computing resources and storage bandwidth.However,in practical applications,the task of fabric defect recognition needs to be completed in real-time on the platform with limited computing power.In order to optimize the depth convolution neural network,this thesis makes use of a kind of advanced depth convolution neural network feature visualization method to carry on the depth feature visualization analysis to the original model.Then,according to the characteristics of fabric images,the network structure is adjusted more pertinently,which not only improves the recognition accuracy of the model,but also reduces the network parameters by more than 90%.A new convolutional neural network LZFNet-V1 is constructed for fabric defect recognition.3)In order to further reduce the computational complexity of deep convolution neural networks,a lightweight convolution structure,which can factorize convolution layers in convolution neural networks,is proposed on the basis of previous studies.Then the standard convolution layer in LZFNet-V1 is replaced with the factorization convolution layer,which can compress the volume of neural network and reduce the calculation consumption of fabric defect recognition system without affecting the recognition accuracy.4)Based on the research of LZFNet-V1,it integrates the most advanced residual mapping and linear bottleneck technology in the field of pattern recognition to form a linear bottleneck convolution module.A new lightweight convolutional neural network LZFNet-V2 for intelligent recognition of fabric defects is constructed by using linear bottleneck module.Compared with the most influential convolution neural network in the world,this model has some advantages in both recognition accuracy and computational efficiency.It is very suitable for fabric defect recognition under the condition of limited computing resources.All the models used in this study can run perfectly in the TensorFlow environment,which makes the lightweight convolution neural network easier to deploy in the industrial field.