基于CNN的火炮身管全景图像疵病识别方法  

Flaw recognition method for gun barrel panoramic images based on convolutional neural network

作　　者：汤一平[1] 韩国栋[1] 鲁少辉[1] 胡克钢 袁公萍 

机构地区：[1]浙江工业大学信息工程学院,杭州310023

出　　处：《仪器仪表学报》2016年第4期871-878,共8页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61070134)项目资助

摘　　要：针对火炮身管内膛疵病种类多、定性定量分析难和检测自动化程度低等问题,本文提出一种以卷积神经网络为基础的疵病识别方法。首先,对全景图像进行预处理,主要包括全景展开、光照强度调整、膛线去除等;其次,通过最优阈值法对图像进行二值化处理,并利用四连通域法提取疵病区域;最后,采用卷积神经网络对疵病进行自动的分类识别。实验结果表明,该方法能有效避免人工疵病特征提取和人工特征描述计算等复杂步骤,实现了"采集-识别-判定"全过程的自动运行,真正实现了窥膛检测的自动化,身管疵病的识别率超过92%,识别准确率远高于基于统计学原理及支持向量机的分类方式,具有较高的准确性,为火炮身管修复及寿命预估等奠定了坚实的基础。Aiming at the problems of variety of bore flaws, difficulty in flaw quantitative and qualitative analysis, and low detection auto- mation level in gun barrel inner flaw detection, this paper put forwards a flaw recognition method based on convolutional neural network （CNN）. Firstly, the panoramic image is preprocessed, including unwrapping the panoramic image, adjusting the illumination intensity, eliminating the rifling, and etc. Secondly, the image is binarized with optimum threshold method, and then the flaws are extracted using 4-connected region method. Finally, the convolutional neural network is adopted to conduct the automatic classification recognition of the flaws. The experiment results demonstrate that the method can effectively avoid the complex procedures, such as artificial flaw feature ex- traction, artificial feature description calculation and etc. , and achieve the whole process automatic operation of ＂acquisition-recognition- judgment＂ in gun barrel inner flaw detection. The automation of the inner bore detection is truly realized. The recognition rate of the gun barrel flaw exceeds 92% , which is superior to those for traditional classification methods based on statistical principle and support vector machine. The proposed method lays a solid foundation for the renovation and longevity prediction of gun barrel.

关 键 词：全景图像 图像处理 卷积神经网络 

分 类 号：TH89[机械工程—仪器科学与技术]