CN-122024337-A - Multi-mode living body detection method

CN122024337ACN 122024337 ACN122024337 ACN 122024337ACN-122024337-A

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a multi-mode living body detection method, which comprises the steps of acquiring two paths of signal related data corresponding to each frame of image of a face of a current detector, calculating pulse intensity sequences of the two paths of signals, performing band-pass filtering and fast Fourier transformation to obtain pulse intensity sequences of the two paths of time domain signals after filtering and main frequencies of the two paths of signals, and calculating physiological pulse correlation intensities; calculating the environment mode credibility according to the gray level change of each mode in each frame of image, fusing the physiological pulsation association strength, the environment mode credibility and the texture probability output by the pre-training convolutional neural network, and determining the living body confidence level. According to the invention, through the synergistic effect of the microscopic physiological characteristics, the macroscopic texture and the environmental perception, the recognition accuracy of the digital counterfeiting attack under the complex illumination environment is effectively improved.

Inventors

WU DONGPENG
Xie Rikai
CHEN KAIWEN
WANG HAO

Assignees

广州技客信息科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A multi-modal living body detection method, comprising: Extracting a forehead region, collecting a visible light signal sequence and an infrared signal sequence of the forehead region, and taking the average value of all signal values in each signal sequence as the pulsation intensity of each signal; filtering the pulse intensity sequences of each signal of all the frame images, obtaining the pulse intensity sequences of each signal after filtering, and performing fast Fourier transform to obtain the main frequency of each signal; Calculating the effective gray scale duty ratio of an infrared mode in each frame of image, constructing dynamic environment weight based on the degree that the average gray scale of the visible light mode in each frame of image deviates from preset ideal imaging brightness, correcting the effective gray scale duty ratio of the infrared mode by using the dynamic environment weight, and determining the reliability of the environment mode; Inputting all frame images into a pre-trained convolutional neural network to obtain texture probability, fusing physiological pulsation association strength, environment modal credibility and the texture probability to determine living body confidence coefficient, and judging living body as living body in response to the living body confidence coefficient being larger than a set threshold value.
2. The method for multi-modal living body detection according to claim 1, wherein the steps of extracting the forehead region and collecting the visible light signal sequence and the infrared signal sequence of the forehead region include: the forehead area is extracted through positioning and identification by a face key point detection algorithm, and visible light signals and infrared signals in the forehead area are respectively acquired through an RGB sensor and an IR sensor.
3. The method of claim 1, wherein the acquiring the pulse intensity sequence of the two filtered signals comprises: The pulse intensity of each signal of all frame images is orderly arranged according to the acquisition time sequence, and is filtered by a fourth-order Butterworth band-pass filter, so that a pulse intensity sequence of each signal after filtering is obtained, and the filtering frequency range is set to be 0.7Hz to 4Hz.
4. The method for multi-modal living body detection according to claim 1, wherein the acquiring the dominant frequency of each signal includes: And converting the pulse intensity sequence of each path of signal after filtering into a frequency domain by adopting fast Fourier transform, and identifying a frequency value of the most concentrated point in the frequency domain space as the main frequency of each path of signal.
5. A multi-modal biopsy method as claimed in claim 1, wherein said determining the physiological pulsation associated intensity comprises: The method comprises the steps of calculating pearson correlation coefficients of pulse intensity sequences of two paths of signals after filtering, calculating the ratio of the absolute difference value of main frequencies of the two paths of signals to the summation result of the main frequencies of the two paths of signals, inputting the ratio as a negative independent variable into a natural exponential function, taking the maximum value of the pearson correlation coefficients and 0, and calculating the product of the maximum value and the output result of the natural exponential function to obtain physiological pulse correlation intensity.
6. A multi-modal biopsy method as claimed in claim 1, the method is characterized in that the determining the credibility of the environment mode comprises the following steps: For each frame of image, the effective gray scale duty ratio of the infrared mode is equal to the ratio of the standard deviation of gray scale values of all pixel points in the infrared mode to the standard deviation reference value of the maximum gray scale calibrated by the infrared sensor; calculating the absolute difference value of the average value of gray values of all pixel points under the visible light mode and preset ideal imaging brightness, summing the preset ideal imaging brightness and the absolute difference value, enabling the dynamic environment weight to be equal to the ratio of the preset ideal imaging brightness to the summation result, multiplying the effective gray duty ratio of the infrared mode by the dynamic environment weight, and taking the average value of the product result of all frame images to obtain the reliability of the environment mode.
7. The method of claim 1, wherein the pre-trained convolutional neural network uses a MobileNetV model.
8. The method of claim 1, wherein determining the confidence level of the living body comprises: dividing the product of the physiological pulsation correlation strength and the environment mode credibility by the square of the physiological pulsation correlation strength, the square of the environment mode credibility and the sum of preset harmonic parameters to obtain a first ratio, and taking the product of the first ratio and the texture probability as the living body credibility.
9. The method of claim 1, further comprising determining a non-living attack and sending an abnormal sample alarm if the living confidence is less than or equal to a set threshold, and preserving current abnormal sample evidence through the interface.
10. A multi-modal biopsy method as claimed in claim 1, characterized in that the method further comprises: and adopting a binocular synchronous camera to acquire images.

Description

Multi-mode living body detection method Technical Field The invention relates to the technical field of image processing. More particularly, the present invention relates to a multi-modal in vivo detection method. Background In the current digital asset authentication scene, the living detection technology is a core barrier for defending identity fraud and protecting digital asset security, and the core aim of living detection is to identify whether a current acquisition object is a real living organism through technical means, but not a non-living organism attack through means such as photos, screen shots or 3D masks and the like, so that personal privacy and digital asset security are seriously threatened, therefore, intelligent detection is required to be carried out on the living organism, thereby effectively defending novel digital counterfeiting attack and ensuring the security of real-name authentication. At present, a mainstream living body detection method in industry mainly depends on single-mode texture analysis or instruction action matching, and the implementation mode of a traditional algorithm generally utilizes Gaussian smoothing filtering to extract edge texture features of an image so as to judge whether reflection features of an unnatural surface exist or not, or judges the effectiveness of biological features by detecting preset actions such as blinking, mouth opening and the like. However, due to the large drastic changes of the environmental light in different authentication scenes and the concealment of the forgery technology, the traditional single-mode or simple fusion technology cannot effectively distinguish real living organisms, for example, images in strong light environments are extremely supersaturated, the traditional fixed logic cannot dynamically adjust the weight emphasis among different modes, and the cooperative evolution relationship of different modes on microscopic physiological characteristics is ignored, so that the digital synthesis trace or frequency abnormality of sub-pixel level cannot be identified, the accuracy rate of judging and identifying the living organisms is greatly reduced when facing the digital forgery attack of highly realistic visual dynamic, and erroneous judgment or omission is easy to generate, thereby seriously threatening the data safety. Disclosure of Invention In order to solve the technical problems that the mode robustness is poor due to the interference of the ambient light and the judgment is inaccurate due to the fact that a fake medium cannot simulate a biological deep microscopic physiological rule, the invention provides a multi-mode living body detection method, which comprises the steps of acquiring each frame of image containing the face of a current detector, wherein the image contains gray information in a visible light mode and an infrared mode; the method comprises the steps of extracting a forehead region, collecting visible light signal sequences and infrared signal sequences of the forehead region, taking the average value of all signal values in each signal sequence as the pulsation intensity of each signal, filtering the pulsation intensity sequences of each signal of all frame images, obtaining the pulsation intensity sequences of each signal after filtering, carrying out fast Fourier transform, obtaining the main frequency of each signal, determining physiological pulsation associated intensity according to the main frequency difference value of the two signals and the correlation coefficient of the pulsation intensity sequences of the two filtered time domain signals, calculating the effective gray scale duty ratio of an infrared mode in each frame image, constructing dynamic environment weight based on the deviation degree of the average gray scale in the visible light mode in each frame image from preset ideal imaging brightness, correcting the effective gray scale duty ratio of the infrared mode by utilizing the dynamic environment weight, determining the reliability of the environment mode, inputting all frame images into a pre-trained convolution neural network to obtain the texture probability, fusing the physiological pulsation associated intensity, the environment reliability and the texture confidence, determining the living body degree, and judging that the living body is larger than a set living body in response to the living body. According to the invention, a closed-loop evaluation system is constructed from two dimensions of microscopic physiological pulsation and macroscopic environment modes by integrating the double-spectrum gray information, the intrinsic consistency is captured by utilizing the physiological pulsation correlation intensity, high simulation digital counterfeiting is effectively identified, the environment mode reliability is introduced, the mode weight is dynamically adjusted, the problem that a single mode is easy to saturate or has large noise under extreme illumination is solved,