Summary of personal work of 21 groups in the first week of summer training

2022.6.21

Attend the summer training opening meeting

• Afternoon 13：30, Participate in the team's first online plenary meeting presided over by the team leader , Mainly specify the contents of the project 、 The main points of 、 technology 、 personnel 、 Division of labor and other issues ;
• Afternoon 15：00, Take part in the Q & a session hosted by Intel's instructors , It mainly introduces the professional content of the project 、 Difficulties and expectations for us , And gave careful answers to our questions .
• evening 19：30, Participate in the first offline plenary meeting of the group , Mainly determine the division of labor , The first milestone of the project is defined ; I am responsible for reading papers related to defect detection of industrial products , Combined with the content of the paper, this paper puts forward targeted suggestions for algorithm optimization .

2022.6.22

• Reading papers ：《 Defect detection methods for industrial products —— review 》, Reading progress is 50%. This paper mainly describes the relevant methods of defect detection of industrial products , For example, the traditional feature-based machine vision algorithm （ Based on texture features 、 Based on color features 、 Based on shape features ）、 Detection method based on deep learning （ Supervision methods 、 Unsupervised methods 、 Weak supervision method ） And possible key problems in defect detection （ Real time questions 、 Small sample problem 、 Small target problem 、 Data imbalance ）. Thesis link https://www.mdpi.com/2076-3417/11/16/7657
• By reading the paper , Combined with the search information , So that I have a basic understanding of industrial product defect detection and related in-depth learning methods , It is equivalent to letting me, who is not familiar with deep learning, enter the door .
• evening 19：30, Participate in online group meetings , The main content is to report the work progress of the first day , Judge whether the division of labor is reasonable , And whether they can complete their tasks before the milestone .

2022.6.23

• Will paper 《 Defect detection methods for industrial products —— review 》 Finish reading , And have a preliminary understanding of deep learning and target detection algorithms .
• Because the team will consider using yolo Target detection , To improve the detection accuracy , So I found an article named 《 Intelligent assembly is based on YOLO v3 Research on industrial part recognition algorithm based on 》 The paper of , And by reading the paper , Trying to find some yolo Optimization method , For after yolo Lay the foundation for the use and optimization of 、 Provide optimization methods . The link to the paper is https://kns.cnki.net/kcms/detail/detail.aspx?dbcode=CJFD&dbname=CJFDLAST2020&filename=GDZJ202010008&uniplatform=NZKPT&v=8PhFCQeAVyTWony9qymh2xyXKaUZFS2BpcRIiVbibehPK7wkPpin_hoVHs5aw8Tz

2022.6.24

• Will paper 《 Intelligent assembly is based on YOLO v3 Research on industrial part recognition algorithm based on 》 Finish reading , I understand YOLO v3 The way to improve ： Use k-means The algorithm re clusters the parameters of the preselection box , Residual network to reduce network parameters , Combined with multi-scale method 、 use Mish Activate functions to improve accuracy , It is more suitable for small target classification detection of industrial parts . The final experimental results show that it is consistent with the original YOLO v3 Algorithm comparison , The improved network model has good robustness , Improved accuracy 1.52%, Improved time 7.25 ms, It realizes the accurate and real-time detection of the parts in the intelligent assembly system .

2022.6.25

• Combined with the introduction of the paper , Mainly learned the relevant knowledge of activation function , Including the definition of the activation function 、 effect 、 Common activation functions （Sigmoid、Tanh、Relu、Mish）、 The formula and image of the activation function , Focus on learning Mish Activation function , Read creation Mish Function of the original paper . come to know YOLO v3 Improved Relu Activation function , Use instead Mish Can improve accuracy , Reduce recognition time .Mish The link address of the original paper ：https://arxiv.org/vc/arxiv/papers/1908/1908.08681v1.pdf
• Try to use Mish Function pair YOLO v3 To optimize
• evening 19：30, Attend group meetings , The main content of the meeting is that each group will report their work of the week . I made PPT, In detail YOLO v3 Optimization method , as well as Mish Function related content .