Research On The Application Of Deep Adversarial Learning In Fabric Image Anomaly Detection

As the world’s largest exporter in the textile industry,controlling the quality of each piece of fabric is a mandatory requirement for the fabric quality inspection in China’s textile industry.At present,most textile companies rely on manual visual inspection or use deep learning models to classify and recognize fabric images to achieve quality inspection,but manual quality inspection is inefficient and prone to misinspection and omission,while most previous deep learning classification models cannot cope with the challenge of unbalanced positive and negative samples of fabric images.Through the study,it is found that the generative adversarial network model can achieve anomaly detection using only normal samples,but the problem of its training instability prevents it from showing the advantages of the model on the detection task.In order to solve the above problems,this thesis firstly implements optimization of the deep convolutional adversarial network-based model,proposes to introduce the ECA efficient channel attention mechanism module under the premise of ensuring stable training of the model,and constructs the ECA-Ano GAN anomaly detection model with better performance for comprehensive detection of fabric defect images.Finally,based on this,a fabric image anomaly detection system based on ECAAno GAN model is designed and developed to help the quality inspection department of textile enterprises to solve practical problems.The research contents and milestones of this thesis are as follows:(1)Construction of fabric defect image dataset.The AITEX defective fabric image dataset is pre-processed,screened and divided to get Base-AITEX dataset and DAAITEX dataset,so that the number and quality of images in the original dataset meet the training requirements.(2)Optimization of the generative adversarial network model and construction of the ECA-Ano GAN fabric defect image anomaly detection model are implemented.To solve the problems of pattern collapse and gradient disappearance of deep convolutional generative adversarial network,this thesis chooses to use Wasserstein distance to replace the calculation of loss function in the original DCGAN network and adds spectral normalization operation to it so that the model can achieve stable training while improving the quality of the model generator and discriminator to achieve the purpose of improving the detection performance of the model.To further improve the model’s attention to the defective region in the detection phase,it is decided to continue to add the ECA attention mechanism module to the discriminator of the optimized model.Finally,the effectiveness of the improvement work in this thesis is verified through experiments,and the final model is compared with other generative adversarial network detection models for experiments,and the experimental results show that the comprehensive detection performance of the model in this thesis is optimal.(3)Design and development of a fabric defect image anomaly detection system.In accordance with the software engineering development method,the system is analyzed for feasibility and requirements,and the system architecture,system function modules and system database are designed in detail.Finalize the development of the fabric defect image anomaly detection system,and ensure the stable and smooth operation of the system through functional testing sessions.