基于改进Logistic回归与不变矩的螺钉滑牙检测方法

Screw sliding detection method based on modified Logistic regression and invariant moment

为检测螺钉滑牙和脱扣等缺陷,搭建测量系统,并对相关算法进行深入研究.根据螺钉对光照的反射特点,选择合适的光源;对采集的图像在二值化基础上进行霍夫圆检测,得到螺钉的位置及大小;对螺钉中心位置进行轮廓提取,并计算Hu矩;将Hu矩作为特征,输入训练好的改进Logistic回归分类器,进行滑牙判断.实验结果表明:在不同环境光照下测量直径5~8 mm的螺钉,采用本方法能很好地检测出是否滑牙,准确率可达95%以上,基本满足对螺钉滑牙检测的要求;同时,系统稳定可靠,有一定的抗干扰能力,当σ=0.01时本文改进Logistic回归分类算法迭代3步后即开始收敛,而传统方法需迭代151步;当σ=0.1时本文方法迭代35步后即开始收敛,而传统方法因学习率过大而发散.

In order to detect the screw sliding and trip,the measurement system is built and the related algorithms used in design of this system are researched.According to the reflection characteristic of the screw,the appropriate light source is selected.The position and size of the screw are obtained by performing the Hough circle on the basis of image binarization.The center of the screw-circle is extracted and the Hu moment is calculated.Hu moments,as characteristics,are input to the improved Logistic regression classifier trained to detect whether the screw is slid-ing.The experimental results show that the screws with diameter of 5-8 mm under different ambient lights can be detected whether the screw is sliding.The accuracy rate can reach more than 95%which meets the require-ments of the screw sliding detection basically,and the system is stable and reliable,and has the certain anti-interference ability.When the learning rateσis 0.01,this modified Logistic regression classification algorithm starts to converge after 33 steps while the traditional method takes 151 steps.Whenσis 0.1,this method starts to converge after 35 steps but the traditional method is divergent due to the excessive learning rate.