基于卷积神经网络的布料疵点检测方法

针对传统布料疵点检测准确率低、识别较慢且计算量大问题，提出基于卷积神经网络的布料疵点检测方法，实现增强布料疵点检测鲁棒性、高效性的设计目标。为保证训练结果准确，首先采集数量以千万级为单位的布料图像并进行图像预处理，标记无疵点布料和疵点布料；然后将图像送入设计的卷积神经网络进行训练和测试，获取疵点检测框；紧接着采用改进的NMS分类算法对检测框进行多框合并，减少误检，进一步提高模型检测效果；最后利用设计的特征图分割算法使网络模型脱离GPU显存限制，适用于各种性能计算机。实验结果表明该方法可在实现布料检测高速度、高准确率的同时增强检测方法的鲁棒性。实际检测速度为3fps，准确率可达99.6%，超过现有疵点检测算法，表明该检测方法可应用于对布料要求更高的生产企业。

Cloth defects Detection based on convolution Neural Network

In order to solve the problems of low accuracy, slow recognition and massive calculation in traditional fabric defect detection, in this paper, a cloth defect detection method based on convolution neural network was proposed to reach the design goal of enhancing its robustness and efficiency. Firstly, in order to ensure the accuracy of the training results, the cloth image in a tens of millions level was collected and preprocessed, and the defect and defect-free cloth images were marked; then the processed image is input into the designed convolution neural network model for training and testing; to be next, the improved NMS classification algorithm is used for multi-box merging to reduce false detection and enhance the detection effect of the model. Finally, the feature graph segmentation algorithm was used to get the deep convolution model out of the limit of GPU memory.The experimental shows that the method can not only achieve high accuracy and high speed of detection, but also enhance the robustness of the method. The actual detection speed is 3fps and its accuracy can be reached at 99.6%, which has exceeded the existing defect detection algorithm, indicating that this method can be applied to manufactures with higher cloth requirements.