基于多因子复杂度的结构误检区域排除算法

在检测金属工件表面细微缺陷过程中,往往将工件的某些结构区域图形误检为缺陷。提出了一种基于多因子复杂度的结构误检区域排除算法;该算法先计算经过卷积神经网络检测之后框选的区域图像的复杂度,即综合计算信息熵,纹理特征及边缘比率的复杂度。根据实验设定合理的复杂度阈值,排除结构误检区域,保留真实缺陷。算法克服了卷积神经网络检测结果中存在一些结构区域的缺点,能够有效去除结构误检区域,并保留非误检区域,有较高的准确率,且计算速度能够达到工业流水线的实时性要求,具有实用价值。

False Detection of Structural Area Exclusion Algorithm Based on Multi-Factor Complexity

In the process of image inspection for slight defects on the metal surfaces, the structural of metal workpiece are often mistaken for some defect regions. In this paper, a structural error detection region exclusion algorithm based on multi factor complexity is proposed. The algorithm first calculates the complexity of the selected region image after the convolution neural network detection, that is, the comprehensive calculation of information entropy, texture features and edge ratio. Set reasonable complexity threshold according to the experiment, eliminate the mistake areas among the detected ones, and keep those real defect regions. This algorithm further remove the false-positive retrieval from the convolutional neural network detection results, can effectively retain the non-structure area error detection area, it has high accuracy, and the calculation speed can reach the real-time requirement of industrial pipeline and has practical value.