基于优化卷积神经网络的表面缺陷检测

Surface defect detection based on optimized convolutional neural network

卷积神经网络(CNN)具有权值数量少,训练速度快等优点,在图像识别、机器视觉等领域得到广泛应用。本文提出了一种卷积神经网络的自适应加权池化算法,算法通过生成合并通道,并在学习掩模的引导下汇集特征,优化了子采样模型的特征提取,有效改善了网络的识别准确性和快速性。利用该算法对磁片表面缺陷进行检测实验,实验结果表明,本文提出的池化模型使卷积神经网络对特征的提取更加精确,同时提高了收敛速度和鲁棒性,并且可以应用于各种深度神经网络体系结构中。

The convolutional neural networks (CNN) have been proved to be effective in image recognition, machine vision and other fields with the advantages of small number of weights and fast training speed. An adaptive weighted pooling algorithm is proposed for convolutional neural networks. The proposed algorithm optimizes the feature extraction of sub-sampling models by generating merge channels and collecting features under the guidance of learning masks, and effectively improves the recognition accuracy and speed of the network. The experiments carried out on the surface defect detection of the magnetic disks show that the proposed pooling model can improve the accuracy of the features extraction and so can effectively detect the defects of surface in faster convergence and robustness with convolutional neural network.Also the pooling model proposed in this paper can be applied to various deep neural network architectures.