Hello, Gattaka, you wrote: G> Thanks for the detailed review. I so understood from the task depends much. I.e. for persons these algorithms, and we tell for search of seals on aerial photographs already others. My task is very similar on car numbers. Here algorithm ? absolutely other algorithms ? to dig for ? Yes, depends on the task much. The method of Violy-Dzhonsa about usage of features of Haar (Haar features) works so well on persons because they have an accurate structure in respect of areas: the forehead is more light, eye-sockets are more dark, the nose too is more light, etc. If we take the task of detection of people (pedestrians) this detector starts to work worse than ever because in a human figure it is more important not than area, and its circuits. About 15 years ago other approach - HOG + linear SVM here "shot". It works actually with circuits (with gradients, it is finite) also not ideally, but it is already much better. Mentioned by me above DPM is more generalized version classical HOG, dlib uses the similar approach for persons and pedestrians, receives better quality at comparable speed of operation. It is called at them FHOG. As classical HOG, and FHOG were checked on car numbers of the small size (from 20 to 100 pixels in height), work very well. It from fast classical methods. Generalization HOG also is ICF (Integral channels features + boosting trees forest) which on the same pedestrians receive better quality, but work noticeably more slowly. Here it is used except HOG also colors (here in what a quality secret), but the approach demands large multisequencing, suits only architecture of type of videocards. There are separate methods for text detection in a frame. They differ from the aforesaid. P.S. But now on are actively used and , Tensor flow not only that under Android so also it is included in basic delivery in its last version. I would not began to sweep aside them if conditions allow. But here not to me, I in the last novelties the small expert.