Search

Optical fiber surface defects have diverse complicated features and different influencing factors. Therefore, the surface defect detection method for optical fiber has good generalization performance. Aiming at the problems of low efficiency, long detection time and high false detection rate in the traditional detection methods of fiber defects on the production line, we establish a database containing three kinds of surface defect samples on the fiber and augmented it in order to reduce over-fitting. This paper proposes a fiber surface detection method based on DSSD algorithm. In the convolutional neural network, the basic network ResNet-101 is utilized to enhance the network feature extraction capability and improve the robustness of the algorithm. The experimental data shows that the detection rate based on DSSD algorithm can reach 96.7%, which proves that the designed fiber intelligent defect detection method can not only greatly reduce the detection time, but also improve the detection efficiency and detection accuracy.

In the visual inspection of flexible materials such as textile surface, the amount of background information is too large, which leads to the problem that the key features need to be extracted is not obvious. The focal loss function is used to construct the RetinaNet and quickly filter out background information, so that the feature information is obtained faster, that is, the surface is flawed. RetinaNet using focal loss is a popular one-stage neural network model in recent years. Based on the feedforward residual network (ResNet), the feature pyramid network (FPN) is used to generate rich multi-scale convolution feature pyramids, which can quickly acquire The characteristics required are important for improving the detection speed of the surface of flexible materials.

For the problem of image distortion in textile flaw detection, a super-resolution image reconstruction technique based on GAN (Generative adversarial network) can reconstruct the obtained low-pixel image into a high-pixel image. The generative adversarial network consists of a discriminative network and a generative network. Generative network is responsible for generate high-resolution images, discriminative network is responsible for identifying the authenticity of the image. the generative loss and discriminative loss continuously optimize the network and guide the generation of high-quality images. The experimental results show that, the PNSR of SRGAN is 0.83 higher than that of the Bilinear, and the SSIM is higher than 0.0819. SRGAN can get a clearer image and reconstruct a richer texture, more high-frequency details, and easier to identify defects, which is important in the flaw detection of fabrics.