This is my problem:
I have a database with face images of various individuals. Each individual have approximately 100 images. In order to recognize some individual by an input image, it's not necessary all 100 images. So, mi team develop a method to select the 5 images that bring better results in future face recognition (not necessarily the 5 images with best quality), in order to reduce the database.
I need papers that talk about methods like this.
As for my opinion this task is seem to be a little bit strange.
1) You may create subspace of indivual images by PCA, for expample. It allow you use maximum information from all images, instead of use infromation directly from five images. There are a lot of papers about it.
2) You may train u own nerual network to recognize every person form your database. There are a lot of papers about it.
3) Moreover, recognition strongly depends from parameters of your current recognition algorithm. You may see parameters influence in almost the same case (face detection) in my papres.
Article Selection of Viola–Jones algorithm parameters for specific conditions
Article Algorithm for optimization of Viola–Jones object detection f...
Conference Paper Some cases of optimization face detection methodes on image ...
Best results come with large databases.
It is not good to reduce datasets but it is possible tu use the adaptive Deep learning (CNN) to get best precision.
See:
https://medium.com/@williamkoehrsen/facial-recognition-using-googles-convolutional-neural-network-5aa752b4240e
Article Robust Face Recognition Using the Deep C2D-CNN Model Based o...
Dear Miguel Alejandro Nicolás Díaz,
look the link, maybe helpful.
https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78
http://www.imedpub.com/articles/a-method-for-face-recognition-from-facial-expression.pdf
Regards, Shafagat
I tried to understand how this problem can be solved by cognitive systems. Models and algorithms are discussed in the book. Prokopchuk Y.A. Sketch of the Formal Theory of Creativity. Monograph. - Dnepr : PSACEA Press, 2017. - 452 p.
(https://www.researchgate.net/publication/319331415_Sketch_of_the_formal_theory_of_creativity?fref=gc&dti=100257367027053 )
Cognitive systems form a "thin slice."
All the best
Maybe, the good solution is to choose faces which have lowest similarity with all other faces (the most original examples). For example, you can match all faces of one person in pairs and then sort them by similarity scores.
Or if your template is n-dimensional vector of float numbers (result of CNN), you can use Principal Component Analysis (PCA), which Alexey Egorov mentioned.
You need a method which turns individual face images (straighten and normalized) to descriptors. Then you can select five descriptors which cover the corresponding vector space. Usually the descriptors are selected by greedy algorithm, so you include the first one, and next one just when the already included descriptors are not sufficient for correct recognition, i.e. there is a descriptor associated to another face which is closer to the the current descriptor than all included ones. This process saturates (under same conditions for image shooting) with small amount of the selected descriptors: four, five or six... (see http://www.agentspace.org/andy/learnopencv). If you need exactly five, then you need more sophisticated algorithm for their selection. Regarding face detection, there is important whether you have front faces only or any faces, which is more complicated. Anyway there is a lot of stuff available just for download: a good resnet model for face detection has come with the latest versions of OpenCV and a good descriptor calculation is provided by resnet model in Dlib. There are also nice e-learning sources for this task: learnopencv.com, pyimagesearch.com.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images