A filter selection method and system that takes both real-time and face quality into account

Abstract

The present invention provides a filter selection method that takes both real-time performance and human face quality into consideration. The filter selection method includes a receiving step, a feature extraction step, a quality scoring step, a quality rating step, a judging step, a storing step and a returning step. The beneficial effects of the present invention are: the present invention adopts supervised machine learning based on real mass data to obtain fuzzy evaluation criteria for the quality of face pictures that are convenient for machines to perform effective retrieval, and quickly decides to select higher quality pictures within a period of time. This method reduces the amount of invalid storage and search under massive data; provides a cascade configuration method for quick filter selection, which achieves both real-time and image quality selection, and also improves flexibility.

Description

technical field [0001] The invention relates to the technical field of face recognition, in particular to a filtering and selection method and system that take into account both real-time performance and face quality. Background technique [0002] Real-time acquisition and selection of high-quality faces from video streams has important practical value. The application of face recognition in real-time streams has great application value in the aspects of case control, arrest, and case warning. However, in real-time video with a large flow of people, the number of faces to be analyzed will increase exponentially, forming a massive analysis source. For a specific face, how to quickly select a higher-quality frame within a certain video time to form a higher-quality effective face data, and how to analyze and store it in a near-real-time manner in a massive face database An urgent problem to be solved in video surveillance system based on face recognition. [0003] The proble...

AI Technical Summary

Problems solved by technology

However, in practical applications, due to the complexity of calculating and comparing each frame of each face in a large-scale, thousands of channels of video one by one, the computing resources are limited, and the time interval in the video is usually selected. This usually results in a large number of collected face pictures, and the pictures collected for the same face will appear repeatedly within a period of time, and the collected pictures may not be correct due to reasons such as angle, size, light, and clarity. Provide high-quality face feature values ​​for search systems based on massive face data, and repeated pictures will also increase additional storage overhead

[0004] The existing technology is based on building a time-series face screenshot linked list for each face in the real-time video stream, which is usually returned to the main application in the form of a callback, but it is not flexible enough in terms of how to filter and make quick decisions

Images