A method for detecting face blur

Abstract

The present invention provides a method for detecting the blurriness of a face. Based on adding blur segments and a hierarchical neural network training method, the aligned face grayscale image, LBP image, and Sobel feature map are used as inputs of a preset neural network model, the face blur category is used as the output of a first neural network, and the face blurriness is used as the output of a second neural network for model training. When the loss function of the preset neural network model tends to be stable, the model training is terminated to obtain the required face blurriness detection model. The present invention can effectively expand the distance between blur segments, improve the accuracy of face blur detection, and especially greatly improve the accuracy of face motion blur detection. It can be convenient to screen out images that meet the needs of face recognition, thereby improving the accuracy of face recognition, and has good practical value.

Description

[Technical field]

[0001] The present invention relates to the technical field of face detection, and in particular to a face blur detection method, in particular a face blur detection method under motion blur. [Background technology]

[0002] Compared with other mature biometric technologies, face recognition technology has been increasingly used in daily life due to its non-contact and strong interactivity. As the input of face recognition, the blurriness of face images will affect the subsequent face recognition process, reduce the final recognition accuracy, and seriously affect the user's face recognition experience in real scenes. Therefore, face blurriness detection is a basic and important process in face recognition technology.

[0003] Face blur can be divided into two categories according to its causes: (1) face defocus blur, which refers to the face blur caused by the camera not focusing on the face, that is, the face is not on the focal plane of the camera; (2) face motion blur, which refers to the face blur caused by the abnormal accumulation of the target face in the horizontal image plane during the exposure time of the camera capture image.

[0004] In the prior art, the detection algorithm for the blurriness of face images is generally based on the gradient operator technology, which characterizes the blurriness of the face by counting the overall gradient value of the face. This algorithm can solve the problem of face defocus blur to a certain extent, but the effect of face motion blur detection is average. Based on this, the present application proposes a face blurriness detection method, especially a detection method under motion blur. [Summary of the invention]

[0005] To solve the above problems, the present invention provides a face blur detection method, which adopts a hierarchical neural network detection method and increases the blur segment of the training data set to effectively improve the detection accuracy of blurred face images, especially the detection rate under motion blur, which can facilitate the subsequent acquisition of images that meet the needs of face recognition, thereby improving the accuracy of face recognition.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a method for detecting the blurriness of a face, especially a method for detecting the blurriness of a face under motion blur, comprising the following steps:

[0007] (1) Obtain multiple face images with different blur levels as training data sets;

[0008] (2) Aligning and graying the face images to obtain an aligned face grayscale image;

[0009] (3) Manually label the aligned face grayscale image and divide it into N categories according to the degree of blur, with the category numbers being 0, 1, ..., N-1;

[0010] (4) The category number is used as an integer and the two decimal places after the random decimal point are used as the blur of the face image;

[0011] (5) Performing traditional LBP processing on the aligned face grayscale image to obtain an LBP image;

[0012] (6) Performing a conventional Sobel filtering operation on the aligned face grayscale image to obtain a Sobel feature map;

[0013] (7) using the aligned face grayscale image, LBP image, and Sobel feature map as inputs of a preset neural network model, using the face blur category number as an intermediate output, and using the face blur as a final output for model training to determine the parameters of the preset neural network model;

[0014] (8) Calculate the loss function of the preset neural network model, and terminate the model training when it tends to be stable, thereby obtaining a face blur detection model;

[0015] (9) Inputting the face image to be tested into the face blur detection model can obtain the blur of the current face image.

[0016] As a technical solution, the face blur category N is preferably 10.

[0017] As a technical solution, the preset neural network model is composed of two levels of neural networks: a first neural network and a second neural network, wherein the first neural network is preferably a MobilefaceNetV2 network, and the second neural network is preferably a BPNet network.

[0018] The present invention effectively expands the distance between blur segments by adding blur segments and a hierarchical neural network training method, thereby improving the accuracy of face blur detection, especially greatly improving the accuracy of face motion blur detection, and has good practical value.

Brief Description of the Drawings

[0019] Figure 1 It is a schematic diagram of the overall flow of a specific embodiment of the present invention.

[0020] Figure 2 It is a schematic diagram of face reference of different blur categories according to a specific embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of the face blur detection process according to a specific embodiment of the present invention. [Specific implementation method]

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment provides a face blur detection method, especially a detection method under motion blur, which specifically includes the following steps: Figure 1As shown:

[0023] S1: Obtain multiple face images with different blur levels as training data sets.

[0024] In this embodiment, the training data set is 195,110 face images with different blurriness from an open source data set;

[0025] S2: Align and grayscale the face images to obtain an aligned face grayscale image;

[0026] S3: Manually annotate the aligned face grayscale images and divide them into 10 categories according to the blur level, ensuring that the number of face images in each category is basically the same. The reference diagram of faces in different blur levels is as follows: Figure 2 As shown, the category numbers from left to right and from top to bottom are 0 to 9, where the smaller the category number, the higher the image blur;

[0027] S4: The category number is used as the integer and the two decimal places after the decimal point are randomly generated as the blur of the face image;

[0028] That is, if the category number of a certain category of face images is i, then the blurriness of each face image in this category is randomly generated between (i).00 and (i).99;

[0029] S5: Perform traditional LBP processing on the aligned face grayscale image to obtain an LBP image;

[0030] S6: Perform a conventional Sobel filtering operation on the aligned face grayscale image to obtain a Sobel feature map;

[0031] S7: Using the MobilefaceNetV2 network as the first neural network and the BPNet network as the second neural network as the backbone network structure to build a preset neural network model, using the aligned face grayscale image, LBP image, and Sobel feature map as inputs of the preset neural network model, using the face blur category as the output of the first neural network and the face blur as the output of the second neural network for model training to determine the parameters of the preset neural network model. The specific network structure is shown in Table 1;

[0032] Table 1

[0033]

[0034] S8: Calculate the loss function of the preset neural network model, and end the model training when it tends to be stable, so as to obtain a face blur detection model;

[0035] S9: Input the face image to be detected into the face blur detection model to obtain the blur of the current face image. The detection process is as follows: Figure 3 shown.

[0036] Based on the win10-Intel(R)Core(TM)i3-8100 CPU@3.60GHz test platform, we used our own test data set containing 29,960 face images for testing, of which 12,000 were out-of-focus blur images, 12,000 were motion blur images, and 5,960 were clear images. The test results are shown in Table 2:

[0037] Table 2

[0038] method Defocus blur detection rate (%) Motion blur detection rate (%) Gradient operator based method 0.89 0.51 The method provided by the present invention 0.99 0.98

[0039] It can be seen from the above table that compared with the traditional detection method based on gradient operator, the method of the present invention has a higher detection rate in both face defocus blur and face motion blur data sets, especially the detection rate of face motion blur has been nearly doubled, effectively improving its use value in practical applications.

[0040] The above embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

Claims

1. A method for detecting blurriness of a face, characterized in that: The steps include: (1) Obtain multiple face images with different blur levels as training data sets; (2) Aligning and graying the face images to obtain an aligned face grayscale image; (3) Manually label the aligned face grayscale image and divide it into N categories according to the degree of blur, with the category numbers being 0, 1, ..., N-1; (4) The category number is used as an integer and the two decimal places after the random decimal point are used as the blur of the face image; (5) Performing traditional LBP processing on the aligned face grayscale image to obtain an LBP image; (6) Performing a conventional Sobel filtering operation on the aligned face grayscale image to obtain a Sobel feature map; (7) using the aligned face grayscale image, LBP image, and Sobel feature map as inputs of a preset neural network model, using the face blur category number as an intermediate output, and using the face blur as a final output for model training to determine the parameters of the preset neural network model; (8) Calculate the loss function of the preset neural network model, and terminate the model training when it tends to be stable, thereby obtaining a face blur detection model; (9) Inputting the face image to be tested into the face blur detection model can obtain the blur of the current face image; The preset neural network model is composed of two levels of neural networks: a first neural network and a second neural network.

2. A face blur detection method as claimed in claim 1, characterized in that: The number of face blur categories N is 10.

3. A face blur detection method as claimed in claim 1, characterized in that: The first neural network is a MobilefaceNetV2 network, and the second neural network is a BPNet network.

Images