Face recognition algorithm evaluation method suitable for public security combats

Abstract

The invention relates to a face recognition algorithm evaluation method suitable for public security combats. The method comprises the following steps of: library establishment test: inputting a sample picture database, establishing a portrait library by a program which carries a to-be-tested algorithm, and acquiring a library establishment speed and a pass rate of the program; recognition test: inputting a plurality of to-be-recognized pictures, selecting a proximate set number of pictures for each to-be-recognized picture from the established portrait library by the program; and result evaluation: obtaining evaluation for the program according to results of the library establishment test and the recognition test. Compared with the prior art, the library establishment process and the recognition process are tested, for query in 1: N test, a plurality of data sets are clearly divided in the recognition process according to data, some data sets are divided according to different photo types, some data sets are divided according to factors which influence the face appearances, and particularly, groups with different age differences are added, so that using performance of old photos in combats is used for reference.

Description

Technical field [0001] The invention relates to an evaluation method, in particular to a face recognition algorithm evaluation method suitable for public security actual combat. Background technique [0002] The face algorithm in the field of public security actual combat mainly has the following application forms: the first type is 1:N comparison, for the unknown identity of the person, query its most similar K person in a given portrait database; the second type is n :N comparison, cross-comparison of two libraries, and judge whether there are the same persons in the two libraries by similarity. In the 1:N comparison, the portrait database is usually very large, such as several million to tens of millions of permanent (or temporary) population. After obtaining the most similar K portrait, the public security personnel can confirm their identity through post-investigation. In the n:N comparison, the purpose is to screen out the criminal suspects whose identity is whitewashed. n...

AI Technical Summary

Problems solved by technology

[0004] 1. The type of comparison does not match: the comparison in academia focuses more on 1:1 comparison, that is, to determine whether two portraits are of the same person. This comparison method is more suitable for civilian scenarios, and is rarely used in public security combat scenarios. use

[0005] 2. The magnitude of the data does not match: the data of the test set in the academic world is small, usually below the order of 10,000, which cannot match the order of tens of millions required by the actual combat of the public security

Experience has shown that when the scale of the portrait library is enlarged by 100 times, the accuracy of 1:N comparison may drop sharply

[0006] 3. Data type mismatch: In the 1:N comparison, although the quality of the query photos may be poor, the large library referred to by N is usually a type of ID photo, which has the characteristics of high definition and uniform posture.

For comparison of portraits with fixed shooting quality like this, the recognition algorithm usually performs specific optimization, and the characteristics of the algorithm cannot be tested if academic data is used

[0007] In addition to academia, various institutions have also conducted face recognition algorithm tests, but it is easy to simplify the test difficulty during the test, thus reducing the reference value for actual combat. For example, the query data all use high-definition recent photos. Easy to collect but will make the algorithm performance much higher

Images