基于编码器解码器的铁氧体零件缺陷检测

Ferrite Parts Defect Detection Based on Encoder-decoder Network

为了提升铁氧体零件缺陷检测的精度和自动化程度,本文使用了一种基于编码器解码器网络的铁氧体零件缺陷检测方法。根据铁氧体零件自身形状特点,设计了一种基于ResNet的编码器解码器网络,通过约800张600×600分辨率的铁氧体零件图像训练该网络,在验证集上像素分类准确率达到99.2%,平均交并比达到86.5%。该模型是一种端到端的网络,在测试阶段,通过模型推理,对零件的缺陷进行定位。该方法的检测精度可以达到5.8μm,满足工业检测的精度要求。

In order to improve the accuracy and automation of ferrite parts defect detection,this paper uses a ferrite parts defect detection method based on encoder and decoder network.According to the shape characteristics of ferrite parts,an encoder and decoder network based on ResNet is designed.The network is trained by about 800 ferrite parts images with 600×600 resolution.The pixel classification accuracy reaches 99.2%and the average cross and union ratio reaches 86.5%on the validation set.The model is an end-to-end network,which locates defects of parts through model reasoning in the test phase.The detection accuracy of the method can reach 5.8microns,which meets the precision requirement of industrial detection.