改进残差网络的轻量级塑料垃圾分类研究

Research on lightweight plastic waste classification based on improved residual network

针对当前日益严重的塑料污染问题,开展塑料垃圾分类与回收迫在眉睫,采用更精细的垃圾分类方法,使得塑料垃圾回收利用的比率将会越来越高,也能从源头上杜绝塑料污染。目前国内塑料垃圾分类网络存在识别准确率不高、运行过程耗时长、模型占用内存大等问题。改进残差网络结构,提出一种双通道轻量级网络(DCLNet-18)对塑料垃圾进行分类。首先使用深度可分离卷积替换标准卷积,减少网络的参数量和计算量;然后在网络的每个BasicBlock后串联双通道注意力机制,增强网络的特征提取能力;最后与AlexNet、VGG、ResNet等系列网络对比。实验结果表明,相比ResNet18网络,DCLNet-18网络的训练时间缩短12.5%,准确率提高2.8%,因此更适合移动端、嵌入式设备应用,为塑料垃圾分类网络轻量化提供了新的思路。

In view of the increasingly serious plastic pollution,it is urgent to launch plastic waste classification and recycling.With more sophisticated waste classification methods,the rate of plastic waste recycling will become higher and higher,and the plastic pollution can be eradicated at the source.At present,the domestic plastic waste classification network has some deficiencies,such as low recognition accuracy,time-consuming running process,and large memory occupation of the model.In this paper,the residual network structure is improved,and a dual-channel lightweight network(dual-channel lightweight network-18,DCLNet-18)is proposed to classify plastic waste.The depthwise separable convolution is used to replace standard convolution to reduce the amount of network parameters and calculations.And then,a dual-channel attention mechanism is connected in series after each BasicBlock of the network to enhance the feature extraction ability of the network.Finally,the proposed network is contrasted with AlexNet,VGG,ResNet and other networks.The experimental results show that,in comparison with the ResNet18 network,the training time of the DCLNet-18 network is shortened by 12.5%,and its accuracy rate is increased by 2.8%.Therefore,it is more suitable for mobile terminals and embedded device applications,and provides a new idea for the network lightweight of plastic waste classification.