基于卷积神经网络的橘子分类识别研究

Research on orange classification and recognition based on convolutional neural network

为了增强橘子表皮缺陷提取效果,满足橘子品质自动分类的实时性和准确性要求,构建了橘子数据集,以ReLU为激活函数,Max-pooling为下采样方法,建立了包含3个卷积层、3个下采样层、1个全连接层和1个Softmax回归分类器为输出层的卷积神经网络模型,采用小批量梯度下降法训练并优化网络模型.实验平台基于Keras深度学习框架,利用Anaconda下的Spyder编译工具进行Python编程,实验结果表明:方法分类准确率达94.34%,比现有分类方法准确率高出4.75个百分点.

In order to enhance the extraction effect of orange epidermis defects,meet the real-time and accurate requirements of automatic classification of orange quality,the orange data set is constructed firstly,then a convolutional neural network model with three convolution layers,three sub-sampling layers,one full connection layer and one Softmax regression classifier is established with ReLU as the activation function and Max-pooling as the sub-sampling method,a small batch gradient descent method is used to train and optimize the network model.The experimental platform is based on the Keras deep learning framework and uses the Spyder compilation tool under Anaconda for Python programming.The experimental results show that the accuracy of the proposed method is as high as 94.34%,the accuracy rate is 4.75 percentage points higher than the existing classification methods.