基于可变形密集卷积神经网络的布匹瑕疵检测

Fabric Defect Detection Based on Deformable Dense Convolutional Neural Network

针对传统布匹瑕疵检测方法无法适用于尺度变化大、面积占比小的瑕疵特征,提出一种基于可变形密集卷积神经网络模型。为了关注到图像中距离较远的特征信息,并避免捕获纹理信息,采用可变形卷积来增强特征的语义表达能力。通过在卷积层中设置卷积像素相对于中心像素各自的x,y方向偏移量,并利用反向传播训练偏移量以增加感受野的变形适应性。同时,采用密集连接的方式以保持模型不遗漏边缘瑕疵信息。最后,根据瑕疵类别预测和位置边框回归实现瑕疵的分类和定位检测。实验结果表明:该模型的平均检测精度和单类目标检测精度标准差分别为93.53%,2.5139,相比于其他方法更具有竞争力。

A deformable dense convolutional neural network is proposed for the traditional fabric defect detection method which cannot be applied to defect features with large-scale changes and small area ratios.To pay more attention to the feature information that is far away in the image and avoid capturing the texture information,deformable convolution is employed to enhance the semantic expression of the feature.By setting the respective direction x and y offsets of the convolution pixels relative to the center pixel in the convolution layer,and using backpropagation training offsets to increase the deformation adaptability of the receptive field.Meanwhile,a dense connection manner is utilized to keep the model from missing edge defect information.Finally,the classification and location detection of defects is realized based on the defect prediction and the border regression.Experimental results show that the average accuracy of the proposed approach and the standard deviation of single-type target detection accuracy is 93.53%and 2.5139,respectively,compared with other methods,it is more competitive.