基于级联式卷积神经网络的芯片检测方法

Chip Recognition and Location Detection Based on Cascaded Convolutional Neural Network

在智能物料排序系统中,芯片存在大小不一、分布密集、摆放角度多样性的情况。针对小体积和带有角度的目标在大尺寸拍摄图像中检测精度较低的问题,提出了一种基于级联式卷积神经网络的芯片识别定位方法来检测芯片的位置和角度。搭建二级级联神经网络,第1级网络利用改进后的YOLOv5s模型检测芯片位置,根据所得位置裁剪出单个芯片图像作为第2级网络的输入,在第2级网络中提出利用关键点回归方法求得角度信息;对大尺寸图像进行叠加式裁剪算法处理,得到尺寸较小的子图作为YOLOv5s模型的输入,解决大尺寸图像对小目标检测精度较低的难题。实验结果表明,基于级联卷积网络的芯片检测精度达94.6%,角度误差小于0.5°,能够满足智能物料排序系统中对位置和角度的要求。

In the intelligent chip auto-layout system,the chips have different sizes,dense distribution,and diverse placement angles.To deal with the problem of low detection accuracy of small-size chips and chips with angles in large-size captured images,a chip recognition and location method based on cascaded convolutional neural networks(CNN)is proposed to detect the position and angle of the chip.In particular,a two-stage cascaded neural network model is proposed.At the first stage,a YOLOv5s model is used to locate the chips.The images of all detected chips based on their bounding boxes are cut and grouped together as the input of the second stage network.For the second stage,it is proposed to use key-point regression to obtain the angles of the chips.In addition,an overlay cropping algorithm is performed in order to partition the large image into multiple images with smaller sizes as the inputs of YOLOv5s,so that each derived image can be processed independently with less computation power.The experimental results show that the accuracy of the proposed chip detection method can reach 94.6%,and the angle error is less than 0.5°,which meets the requirements of intelligent chip auto-layout system.