基于深度信念网络的织物疵点检测

Fabric defect detection based on deep-belief network

为提高织物疵点检测精度和效率,提出了一种基于深度信念网络的织物疵点检测方法。用改进的受限玻尔兹曼机模型对深度信念网络进行训练,完成模型识别参数的构建。利用同态滤波方法对图像进行预处理,使疵点图像更加清晰,同时抑制了背景图像。以Python语言,基于TensorFlow框架构建深度信念网络模型,对织物疵点图像进行处理得到学习样本,确定模型激活函数后,分析了各模型参数对织物疵点检测准确率的影响规律,得到激活函数为Relu, Dropout值为0.3,预训练学习率为0.1,微调学习率为0.000 1,批训练个数为64时,模型参数值达到最优。最后,利用在无缝内衣机上采集到的各类疵点图像,对深度信念网络织物疵点检测模型进行验证。结果表明:所提出的织物疵点检测方法能够快速、有效地对织物疵点进行检测和分类识别,准确率达到98%。

Objective In order to improve the quality of textile products, increase economic benefits, and reduce production costs in order to improve production efficiency, it is of great significance to achieve intelligent detection of fabric defects. A fabric defect detection method based on deep-belief network(DBN) is proposed, and the deep-belief network is trained by the improved restricted Boltzmann machine model to complete the construction of model recognition parameters, which can not only independently extract fabric image data features, screen effective information for transmission, but also have a short training time and fast model convergence speed.Method In order to make the best training effect of the model, this paper enriches the samples by using the data augmentation method to meet the training requirements of the DBN model. The homomorphic filtering method is used to preprocess the image to reduce the low frequency and increase the high frequency, and sharpen the details of the image edges, making the defective image clearer and suppressing the background image. In order to solve the overfitting problem of the model and improve the generalization ability of the model, the DBN-Dropout model is used to set the output information in the network to 0, the contrastive divergence method is employed to initialize the visual layer of the training sample, the activation probability of neurons in the hidden layer of the model is calculated, and the activation status of neurons in the hidden layer and the visual layer is assessed. In Python language, a DBN model is built based on the TensorFlow framework, and the learning samples are obtained by processing the fabric defect images. In the weft knitting laboratory of the Knitting Engineering Technology Research Center at Zhejiang Sci-Tech University, the area scan CCD camera was combined with the 6 mm focal length lens(FL-HC0614-2M) produced by Ricoh Corporation of Japan, and 200 images were collected of different types of plain weft knitted fabrics produced by the RFSM20 high-speed seamless underwear machine developed by Zhejiang Rifa Textile Machinery Co., Ltd., including 50 images of normal flawless fabrics and 150 images of fabrics with various defects. The collected fabric image samples were grayscale images with a size of 512 pixels×512 pixels, and the 100 fabric images collected were detected by the DBN model.Results Cross-entropy and Adam were used as loss functions and optimizers respectively, the activation function, loss function and optimizer of the model were studied with the fabric defect dataset, and then the activation function, Dropout value, learning rate and training batch number of the model were analyzed with the fabric defect dataset. The activation function was a Relu function, whose Dropout value was 0.3, pre-training learning rate was 0.1, fine-tuning learning rate was 0.000 1, and batch training number was 64, model and the parameter values were optimal. 512 pixel×512 pixel fabric images were used to conduct experiments in MatLab2019b environment, and the results of the experiments on the TILDA dataset using particle swarm optimization(PSO)-BP neural network and local contrast deviation method were compared with the algorithm proposed in this paper, and it is concluded that the proposed deep-belief network has better detection results and clearer target contour recognition for fabric defect detection under complex background than the other two methods. In order to evaluate the applicability of this algorithm, defect images weft knitted fabrics with different types of defects were used for detection. For the practical testing of 100 pictures, the experimental results were combined with the detection effect chart, which showed that the algorithm introduced in this paper has good detection results for fabric holes, oil pollution and yarn breaking defects, and its detection accuracy rate reaches 98%, which verifies the algorithm′s adaptability, effectiveness and accuracy for fabric defect detection.Conclusion The experimental results show that the DBN network model has a good detection effect on fabric defects, which can not only identify the shape and outline of fabric defects, but also detect different types of fabric defects, which shows the effectiveness of the algorithm.