Fabric Defect Detection Method Based on Improved U-Net

Abstract

Computer vision builds a connection between image processing and industrials, bringing modern perception to the automated industrials. At the same time, defect detection based on deep learning has played an important role in automated detection. In this paper, an improved convolutional neural network CU-Net for fabric defect detection is proposed. In this method, the classical U-Net network was improved. On the basis of network size compression, attention mechanism is introduced and a new compound loss function is used for training. Using the public AITEX defect fabric data set as the test sample, the experimental result shows that the accuracy and recall of the proposed method are 98.3% and 92.7%, respectively. Compared with the highest scores of other detection methods, they are improved by 4.8% and 2.3%, which improves the detection accuracy of fabric defect significantly.

Ensemble learning application for textile defect detection

“…The publicly available AITEX defective fabric dataset was used as the test dataset. When the results of the study were examined, it was seen that the proposed method had an accuracy of 98.3% [13].…”

Defect detection in textile manufacturing: Improving findings using YOLOv5 versions

“…Apesar dos demais resultados e da otimizac ¸ão do recall, com um resultado de 83,82%, foi possível superar apenas [17] (82,58%), permanecendo abaixo dos demais trabalhos que utilizaram outros métodos não relacionados ao YOLO na base de dados AITEX. No topo, [20] alcanc ¸ou um recall de 92,70%. Mesmo não superando o melhor resultado, o resultado obtido ainda está muito acima dos demais trabalhos que utilizaram o YOLOv5, como [18](48,95%).…”

Textile defect detection using YOLOv5 on AITEX Dataset

“…Com isso, alcanc ¸aram resultados superiores ao modelo U-Net com menos recursos computacionais e menos parâmetros, e consequentemente menor tempo. Na mesma linha de pesquisa, [Rong-qiang et al 2021] construíram uma versão aprimorada da rede neural U-Net reduzindo o número de canais e adicionando um mecanismo de atenc ¸ão com quatro vezes mais características que a original para aumentar a precisão da localizac ¸ão. Além disso a detecc ¸ão de defeitos foi tratado como um problema de classificac ¸ão binária, identificando se os pixels pertencem ou não à classe defeituosa.…”