Research On Chip Surface Defect Detection Algorithm Based On Deep Learning

Today,with the rapid development of application fields such as industrial control,electronic products,medical devices,and communication engineering,the quality requirements and demand for chip production continue to increase.Chip defect detection,as an indispensable part of the chip process,affects the quality and performance of chip products,and is particularly important today.Chip defect detection can actively feedback the production quality of the previous process and promote the improvement of chip production level.Traditional chip defect detection mainly uses manual visual inspection methods,which have the disadvantages of strong subjectivity,high training costs,and low detection efficiency.Therefore,designing high-precision and high-efficiency chip defect detection algorithms has important application significance and value.Therefore,this article combines target detection technology with the field of chip defect detection to study an improved chip defect detection algorithm based on YOLOv5 s.The main work of this article is as follows:Firstly,Based on the chip defect images collected in the actual production process,sample screening is completed,chip defect datasets are constructed,and artificial image annotation is performed.For the imbalance between classes in the original chip defect datasets,data enhancement strategies are developed.For small target defects such as scratches and foreign objects,a Copy-Paste based annotation image combination enhancement method is proposed to improve data diversity.After data enhancement,4623 defect instances were ultimately obtained.Secondly,In order to further solve the problems of large network parameters and difficulty in deployment,a lightweight performance optimization study was conducted on YOLOv5 s.This article proposes two model improvement methods based on YOLOv5 s.The first method is to optimize the feature extraction of the backbone network based on the Ghost Net module.GSConv is used to replace the conventional convolution of the neck to reduce the parameter amount.The improved model improves by 1.1% compared to the benchmark model m AP,and the calculation amount and model volume are reduced by 55.06%and 36.64%,respectively.At the same time,the detection speed reaches 68.49 FPS.The improved model is named YOLO-CDD-L model.The second is a trunk feature extraction network optimized based on Shuffle Netv2 module,and an improved GSConv neck convolution model.Aiming at the problem of reduced detection accuracy after the lightweight improvement of the model,an improvement based on the CA attention mechanism has been made.The improved model,while ensuring a certain detection accuracy,has reduced computational complexity and model volume by 89.24% and 87.59%,respectively,and increased detection speed by 10.9 FPS,Name the improved model YOLO-CDD-S.The experimental results show that the YOLO-CDD-L model and YOLO-CDD-S model proposed in this paper significantly reduce the computational complexity and model volume of the model while ensuring the application performance of the model,verifying the effectiveness of the improved method and meeting the real-time detection requirements of edge devices and mobile terminals.Finally,An easy-to-use chip defect detection system platform was built using Py QT5,and results were visually analyzed based on chip defect data sets.The design and implementation of a chip defect detection system has the characteristics of simple operation and friendly interface,and is tested based on chip defect data sets.The experimental results show that the system can accurately and quickly detect chip defects.