****************************************************************************
In this paper, facial image recognition AI comparison test
1. Testing Requirements Analysis
2. Test preparation environment
3. Test data preparation
4. Test analysis and execution
5. Test Summary of issues
****************************************************************************
First, the needs analysis
1) functional requirements
AI automatic attribution photo (off-line model): Identify the principal, teachers, parents, relatives and friends in the growth time, class rings, parent-child activities inside the uploaded photos, and alignment with human face reference library, get a picture of where the unique identification and It contains the child's unique identification (group photo may contain more than one child) finally generated growth of young children archives
2) test requirements point:
Face reference library function test
Face matching function / model assessment test
Automatic attribution integration tests (test network now scheduled task)
3) than the picture of the flow chart
Second, the test environment preparation (installation dependencies, TensorFlow library solve the problem)
1) install dependencies (dependencies 17)
2) dependent on ambient Anaconda installation TensorFlow library, see below link:
https://www.cnblogs.com/xjx767361314/p/11103817.html
Third, (algorithm engineering Results and discussion) to prepare test data
1) data collection
Children: 3--6 years old baby
Photo format: png, jpg
Image requirements: a reference library (the best according to a front upper body) daily picture (about 5)
Pictures of key information: front, side, single photo, group photo
2) child care picture data recording label
Data 3) children labeled
Fourth, test execution and results statistics
1) Analysis Test procedure: first generating a reference library feature vector children and different types of input images to be recognized are denoted verification image data according to the comparison result, the results of the statistical evaluation than the effect output model
2) execute the script, a picture comparing children
3) Analysis of the recognition result (the manual way than the execution result, efficiency is very low)
4) the results of statistical comparison
Fifth, test summary problem
1) testing requirements phase is not explicitly model the effect, only the collective model according to photos and blur effect is not ideal situation is optimized test is not rigorous.
2) test environment preparation takes four days, four days before the test time, the main problem is the first installation of AI picture identification direction dependencies, tensorflow installation error, finally found incompatible with the underlying glibc library, followed by the installation package of nearly 17 , mutual dependencies, the use of domestic mirroring speeds up the installation.
Miniature effect 3) existing network is not very good, mainly because the model has not been trained data actual scene, in fact, insufficient test data, but also with the existing network data there may be differences in the collection of the test set may be too ideal, the actual picture quality child care It is not high, after testing should try to collect data closer to the actual scene.
4) there is no specific test data marker data labeling, testers who manually tag data, if a large amount of test data is bound to increase the cost of time, still looking for a good way.
5) Test results lower than efficiency, it is now one by one comparison and analysis based on tagging data and test results data, as far as possible to automate check Depending on the project, automation output statistics.