2022 Summer Project Training (II)

One 、 Preliminary work arrangement

6 month 21 After holding a group meeting on the th to arrange the division of labor , Considering that there is a certain difference between your knowledge reserve and your understanding of the project , To guarantee 6 month 25 Can successfully complete the planned task on the day , The group is in 6 month 22 The group meeting was held again on the evening of the th , Each member reports his / her work progress of the day , To summarize . After the meeting, I thought the task arrangement was reasonable , Each group can complete the task within the specified time , Work began to advance in an orderly manner .

Two 、 Current stage progress

Thesis group & The project team ：
Transfer learning is used in model training , At the same time yolo Learned .
Analysis of the yolo v1,v2,v3 The structure and characteristics of the algorithm , Can be found v3 Better . stay 《 Intelligent assembly is based on yolo v3 Research on industrial part recognition algorithm based on 》 I learned that mish Activation function , To improve yolo v3 The model can improve the recognition accuracy .
For the problem of less data sets , Propose several data enhancement methods ： rotate 、 The zoom 、 Pan to generate a new picture ; Image features are extracted by Gaussian blur and edge detection . A brief understanding of edge detection Canny Algorithm .

Applet group ：
At present, it is possible to select pictures by applet 、 Send pictures to WEB Server 、 from WEB Get the processing results on the server 、 Display the results , But not yet connected to the back end .

Model optimization group ：
According to the official ppt And documentation steps , Understand the basic api Use steps , Successfully run the official model demo.

Algorithm group ：
The algorithm group works the most , At the same time, it is also responsible for the overall planning and division of labor arrangement of the whole team . The work of the algorithm group covers web front end , Paper reading , Data preprocessing , Data to enhance , Model selection , Model training and other aspects . At present, the reasoning accuracy and speed of the model are better , The next step is to deploy to the backend , At the same time, continue to find solutions to improve the accuracy of the model .

3、 ... and 、 The goal for the next phase

Next stage refers to 6 month 27 A period of time before the mid-term defense , The following tasks are scheduled to be completed before the mid-term defense , And record the specific work content in the mid-term defense ppt.

Thesis group ：
Learn relevant methods of Feature Engineering .

The project team ：
Sort out recent meeting minutes and development logs ; Draw charts according to the needs of other groups .

Model optimization group ：
Study 、 Deploy TF-Serving; Try to use Intel Optimize existing models .

Applet group ：
Connect back end , Upload pictures 、 Test results and time-consuming feedback and display .

Algorithm group ：
Implement a defect location algorithm with confidence ; Carry out certain characteristic Engineering ; Before and after development , Realize the detection of single and multiple pictures ; Use Tensorflow Refactoring some modules .