基于卷积神经网络的工件识别算法

A Work Piece Identification Algorithm Based on Convolutional Neural Network

针对传统工件识别算法特征提取困难、通用性差、工件的平移、旋转和光照变化对识别效果影响较大、识别准确率不高等问题,提出了一种基于卷积神经网络的工件识别算法。卷积神经网络由4层网络构成,包括2层卷积层和2层全连接层。实验任意选取了10种工件进行识别。在神经网络训练阶段对这10种工件共采集1万张图片,其中9000张图片作为训练集,剩下1000张图片作为验证集。训练时采用在卷积层加入批归一化层和在全连接层使用随机失活的方法,使网络能够得到更好的训练效果。当迭代次数达到10万次时基本得到理想的训练效果。测试时通过摄像机采集图像,对采集到的图像进行预处理,然后将预处理后的图像送入网络进行识别。在光源稳定室内环境下进行实验,实验结果表明基于卷积神经网络的工件识别平均所需时间为0.169s,平均识别准确率为98.3%,准确率高于传统基于特征提取和模板匹配的工件识别。

A work piece identification algorithm based on convolutional neural network is presented to solve the problem of traditional work piece recognition algorithm,which is the features hard to extract,the universality is poor,the shift and rotation of work piece and the change of light have great influence on recognition.Convolution neural network consists of four layers of network,including two layers of convolution and two layers of full connection.The ten kinds of work pieces need to be identified were selected randomly in the experiment.A total of 10000 pictures were collected in the training stage,of which 9000 images were used as a training set,and the remaining 1000 pictures were taken as a verification set.By adding the normalization layer in the convolution layer and using dropout in the full connection layer to make the network get better training results.When the number of iterations reached 100000,the ideal training effect resulted.In the testing stage,the images which collected by the camera are preprocessed and are inputted into the network for identification.The experimental results show that the average time required for work piece identification based on convolutional neural network is 0.169s,and the average recognition accuracy is 98.3%,which proves that accuracy of the algorithm is higher than that of traditional work piece identification algorithm based on feature extraction and template matching.