卷积神经网络的贴片电阻识别应用

Chip resistance recognition based on convolution neural network

贴片电阻由于其体积微小、性能稳定等独特的性质,在当今智能化的电子设备中被广泛使用。为保证贴片电阻的出厂质量,需要对其进行缺陷识别、极性方向识别、正反面识别和种类识别,目前很大程度上依靠人工肉眼进行识别检测,效率低、容易误检、成本高。本文针对传统图像识别方法的局限性,结合近年来卷积神经网络在图像识别方面所取得的巨大成就,基于AlexNet模型、GoogLeNet模型、ResNet模型思想设计了3种深度适宜、可训练参数约4×10~6(百万)的卷积神经网络,克服了当前主流卷积神经网络模型由于可训练参数过多、模型层数太深导致在贴片电阻识别应用中识别速度不能满足实时性要求、泛化识别准确率低的问题。实验表明,3种模型的识别准确率均超过90%,最高识别准确率达到95%,识别速度达到0.203 s/张(256像素×256像素,CORE I5)。因此,本文设计的3种卷积神经网络可根据具体实际需求进行选用,在实践中具有极强的可行性和可推广性,同时也在提升企业生产效率和产品质量方面具有重要意义。

Chip resistors are widely used in intelligent electronic devices because of their unique properties such as small size and stable performance.The chip resistors produced by the factory must be identified for defects in both front and back faces,polarity,and type in order to guarantee the quality.However,such identification largely relies on the eye detection,which is inefficient,prone to error,and costly.In this paper,considering the limitation of the traditional image recognition methods and the great achievements of convolutional neural network(CNN)in image recognition in recent years,three CNN models,AlexNet model,GoogLeNet model,and ResNet model,with appropriate depth and training parameters of about 4 M(million)are designed to overcome the demerits of low speed that results in the inability to meet the real-time requirement.These models overcome the low accuracy problem of generalization recognition associated with the prevailing CNN models,which is caused by many trainable parameters and many layers of model.Experiments show that the recognition accuracy of these three models exceeds 90%.The highest recognition accuracy rate is 95%,and the recognition speed is 0.203 s/piece(256×256 pixels,CORE I5).Therefore,these three CNN models can be adopted in practice and have a strong feasibility and replicability;thus,they have a great potential to improve the production efficiency and product quality for chip resistors.