Abstract: To address the problem of the appliance recognition model being interfered by different classes during the recycling process of waste appliances,causing unstable recognition results,an incremental learning with dynamic prototype for waste appliance recognition was proposed.An incremental residual aggregation structure was established to acquire new and old class appliance features to enhance the extension capability of the appliance recognition model.A shared weight dynamic prototype was designed to acquire representative features and distinguishing features of appliances,reducing the cross-interference in the recognition process.Finally,a comparison prototype method was designed to perceive the misclassification,combining the appliance representative features of the shared weight dynamic prototype to improve the recognition accuracy.The proposed recognition method was applied to the sorting of waste appliances in different scenarios,and the experimental results demonstrated the recognition accuracy of the method.

Key words: waste electrical appliance recognition, dynamic prototype, incremental learning, deep neural networkwaste electrical appliance recognition, dynamic prototype, incremental learning, deep neural network

摘要： 针对废旧家电回收过程中废旧家电识别模型受到不同类别干扰,引起识别结果不稳定的问题,提出了一种基于动态原型增量学习的废旧家电识别方法。首先,建立增量残差聚合结构,获取新旧类家电特征,增强了废旧家电识别模型的扩展能力。其次,设计共享权重动态原型,获取家电代表性特征和区分性特征,降低了识别过程的交叉干扰。最后,设计对比原型方法感知误分类别,结合共享权重动态原型的家电代表性特征,提升了识别精度。将提出的识别方法应用于不同场景下废旧家电分拣,实验结果表明该方法具有较好的识别精度。

关键词: 废旧家电识别, 动态原型, 增量学习, 深度神经网络