Defect automatic diagnosis and repair method and device for convolutional neural classification network

The invention discloses an automatic defect diagnosis and repair method and device for a convolutional neural classification network. The method comprises the following steps: 1) predicting importance normalization of the convolutional neural classification network; 2) calculating importance scores of the activation graphs; 3) performing regular statistics on a normal sample activation graph; 4) classification error sample activation graph screening; the full-automatic defect diagnosis and repair method for the convolutional neural classification network is designed and used for defect detection and automatic repair of the pre-trained convolutional neural classification network model, and the classification performance of the deep convolutional neural classification network model can be effectively improved. 本发明面向卷积神经分类网络的缺陷自动诊断与修复方法及装置，包括下列步骤：1）卷积神经分类网络预测重要性归一化；2）激活图重要性得分计算；3）正常样本激活图规律统计；4）分类错误样本激活图筛选；5）卷积神经分类网络自动化缺陷修复。本发明设计的是一种面向卷积神经分类网络的全自动缺陷诊断与修复方法，用于已经预训练的卷积神经分类网络模型缺陷检测与自动修复，能够有效提升深度卷积神经分类网络模型的分类性能。