【摘要】 为了在硬件资源受限的设备上实现通过木材切面图像自动识别木材种类,提出一种轻量化卷积神经网络。此轻量化卷积神经网络是在MobilenetV2的基础上,去除了部分多余的反向残差块,降低了反向残差块的通道扩张系数,从而大幅降低了计算量和参数量。为了提高网络的泛化性能,此轻量化卷积神经网络训练时采用标签平滑策略。实验结果表明,该网络与深度学习常用模型和传统机器学习方法相比,识别率更高,达到了99.22%,占用存储空间少,计算效率高。更多还原

【Abstract】 To realize the automatic identification of wood species through wood section images on devices with limited hardware resources, a lightweight convolutional neural network is proposed. Based on MobileNetV2, this lightweight convolutional neural network removes some redundant inverted residual blocks and reduces the channel expansion coefficient of the inverted residual blocks, thus greatly reduces the amount of calculation and parameters. To improve the generalization performance of the network, this lightweight convolutional neural network adopts a label smoothing strategy during training. The experimental results show that the network has a higher recognition rate compared to current deep learning commonly used models and traditional machine learning methods, reaching 99.22%, consuming less storage space, and high computing efficiency.更多还原