基于改进的卷积神经网络的工件识别技术

Workpiece recognition technology based on improved convolution neural network

针对传统的工件识别流程是由人工根据工件的特点设计需要提取的特征,整个过程具有耗时高、成本大、通用性较差和识别准确率不高等问题,改进了经典的卷积神经网络模型AlexNet和LeNet-5,通过将AlexNet网络的输入图像尺寸缩小到120×120,用BN层代替LRN层,减少两层卷积和全连接层,用3×3的卷积核代替第一层卷积层11×11的卷积核;将LeNet-5的输入图像尺寸提升至60×60,用ReLU取代原始Sigmoid激活函数,使用多个小卷积核代替大卷积核;分别使用改进前后的网络模型对工件数据集进行训练、测试.结果表明,改进后的两种网络模型,在测试集上分别达到94.31%和92.75%的平均识别准确率,平均识别时间分别为0.271s和0.321 s,满足生产需求.

In response to the traditional workpiece recognition process in which the features to be extracted were designed manually according to the characteristics of the workpiece,the whole process has the problems of high time consumption,high cost,poor generality and low recognition accuracy,the classical convolutional neural network models AlexNet and LeNet-5 were improved by reducing the input image size of AlexNet network to 120×120,replacing the BN layer with By reducing the input image size of AlexNet to 120×120,replacing the LRN layer with BN layer,reducing two convolutional and fully connected layers,and replacing the 11×11 convolutional kernels of the first convolutional layer with 3×3 convolutional kernels;increasing the input image size of LeNet-5 to 60×60,replacing the original Sigmoid activation function with ReLU,and using multiple small convolutional kernels instead of large convolutional kernels;and using the improved network models before and after training and testing on the artefact dataset,respectively.The results showed that the two improved network models achieved an average recognition accuracy of 94.31%and 92.75%respectively on the test set,with an average recognition time of 0.271s and 0.321 s respectively,meeting the production requirements.