【摘要】 为了解决游标卡尺检定系统自动读数这一关键问题,提出一种基于灰度交替投影的数字区域分割和基于卷积神经网络的数字识别方法。首先根据图像中主标尺、游标尺和数字区域的空间位置及灰度分布特点,多次交替对图像进行灰度水平投影和竖直投影,进而分割出卡尺数字识别的感兴趣区域。然后,通过卷积神经网络的卷积层提取卡尺数字的图像特征,利用池化层降低特征维度,最后由全连接层和softmax激活函数输出数字分类结果。实验结果表明,交替投影法的误分割率小于1%,低于最大连通区域法与连通区域重心法,卷积神经网络的数字识别正确率高达99.5%,精度波动只有0.05%,明显优于线性分类器和KNN算法。本文方法很好地实现了工业用游标卡尺自动检定过程中的数字识别。更多还原

【Abstract】 To solve the problem of automatic number-reading in the vernier caliper verification system,region segmentation method based on alternate gray projection and number-recognition method based on convolutional neural network are proposed. Firstly,according to the spatial position and gray distribution characteristics of the main ruler,vernier ruler and numbers,the image is alternately projected horizontally and vertically for many times,and then the region of interest for numbers is segmented. Then,the image features of numbers are extracted by convolutional layers of the neural network,and the feature dimension is reduced by pooling layers. Finally,the numbers classification results are output with fully connected layer,which activation function is softmax. The experimental results show that the error segmentation rate of alternate projection is less than 1%,which is lower than the methods of both maximum connected region and gravity center of connected region. The accuracy of convolutional neural network is 99.5% with fluctuation only 0.05%,which is obviously better than the linear classifier and KNN algorithm. The method has realized the digital recognition in the automatic verification of vernier calipers in industry.更多还原