基于卷积神经网络的废旧塑料瓶颜色分拣系统

Intelligent Sorting Method of Plastic Bottles Based on Convolution Neural Network

针对分拣环境中塑料瓶颜色变化多样、依靠经验设计的特征提取方法鲁棒性不强的问题,提出一种基于卷积神经网络的废旧塑料瓶颜色分类识别方法。首先收集塑料瓶并采集图像进行人工标注制成4类塑料瓶图像数据集,然后设计具有残差连接的卷积神经网络,引入数据增广技术、真实标签平滑处理和在全连接层中引入具有自归一化性质的SeLU激励函数等措施对其加以训练。实验结果表明,所提出的方法准确率高达99.2%,优于传统方法,并且能够满足实时性要求。

To cope with the problems in the waste plastic bottle sorting,in terms of diverse types and the low level of robustness of the feature extraction method,this work developed a recognizing and categorizing method of waste plastic bottles based on a convolution neural network(CNN).The plastic bottles are firstly picturized and categorized into 4 groups of data.Then,a convolutional neural network with residual connections has been designed.The model is trained based on augmentation technology,real label smoothing,and other regularization measures.Experimental results demonstrate that the accuracy of the developed method is up to 99.2%,which is better than the traditional method and meets the real-time requirements.