注意力可变形卷积网络的木质板材瑕疵识别

Attention Deformable Convolutional Networks for Wooden Panel Defect Recognition

为了解决木材缺陷检测中人工成本高、效率低的问题,该文基于可变性卷积网络和注意力机制,提出一种端到端的神经架构模型.首先,可变形卷积网络(Deformable Convolutional Network, DCN)通过将矩形网格转换为变形网格,使模型专注于具有更多有用图像信息的区域.使用可变形卷积网络可以忽略图像特征中不相关的系数,解决了传统卷积在特征中学习更多信息能力有限的问题.然后,将DCN输出馈送到门控循环单元(Gated Recurrent Unit, GRU)层以学习缺陷图像的高级特征.最后,通过关注输入图像的最重要特征,应用注意力机制加强瑕疵区域的高亮度,从而提高模型识别的准确性.使用Matlab平台在4个木质板材缺陷数据集上将该文方法与现有其他方法进行比较分析,该文方法的准确率比其他3种对比方法提高了2.4%~13.2%的维度,灵敏度提高了3.3%~16.6%的维度,特异性提高了4%~21%的维度.实验结果表明,该文方法在检测精度和其他各个性能方面均优于现有方法,最佳准确率为99.2%,证明了该文方法的有效性.

To solve the problem of high labor cost and low efficiency in wood defect detection,this paper proposes an end-to-end neural architecture model based on a deformable convolutional network and an attention mechanism.Firstly,the deformable convolutional network(DCN)enables the model to focus on regions with more useful image information by converting a rectangular grid into a deformed grid.Using a deformable convolutional network can ignore the irrelevant coefficients in image features,addressing the limited ability of traditional convolution to learn more information in features.Then,the DCN output is fed to the gated recurrent unit(GRU)layer to learn high-level features of the defective image.Finally,by focusing on the most important features of the input image,the attention mechanism is applied to enhance the high luminance of the defective regions,thus improving the accuracy of the model recognition.Using the Matlab platform to compare and analyze this paper's method with other existing methods with four wood panel defect datasets,the proposed method improved the accuracy by 2.4%to 13.2%in dimensions,sensitivity by 3.3%to 16.6%in dimensions,and specificity by 4%to 21%in dimensions over the other three compared methods.The experimental results show that the method in this paper outperformed the existing methods in terms of detection accuracy and various other performance aspects,and the best accuracy was 99.2%,which proves the effectiveness of the proposed method.