管道漏磁图像的卷积核信息熵相似度约束方法

为了提高卷积神经网络中卷积核对管道漏磁图像的特征学习能力,提出一种基于信息熵相似度约束的卷积核优化方法.建立一种信息熵相似度约束规则,通过判定条件对权值相近或相似度过高的卷积核进行优化.搭建实验平台并建立样本数据库进行实验,该方法可优化其特征提取能力,得到具有区分目标和背景语义信息能力的优化卷积核.结果表明,优化卷积核对目标具有较强的特征提取能力,能提高网络分类准确率和工作效率,实验结果与理论分析具有很好的一致性.

Constraint method for convolution kernel information entropy similarity of pipeline magnetic flux leakage images

In order to improve the feature learning ability of convolution kernel for pipeline magnetic flux leakage images in convolutional neural networks(CNN), a convolution kernel optimization method based on information entropy similarity constraint was proposed. An information entropy similarity constraint rule was established, and the convolution kernel with similar weights or extreme similarities was optimized by judging conditions. An experimental platform and a sample database were built for experimentations. The as-proposed method could optimize the feature extraction ability and get the optimized convolution kernel capable of distinguishing target and background semantic information. The results show that the optimized convolution kernel has strong feature extraction ability for the target, and can improve the accuracy and efficiency of network classification. The experimental results are in good agreement with the theoretical analysis.