Face feature extraction method, low-resolution face recognition method and device

Abstract

The invention discloses a face feature extraction method, which comprises the steps of collecting a face image and a feature extraction network, extracting features by using an initial feature extraction module and a GTFB module, fusing the features by using a bottleneck layer, amplifying a feature map, performing feature transformation by using a feature transformation sub-network and the like. The amplification subnet and the feature transformation subnet are connected to form a network, and redundant useless information is reduced by reusing important information on different scales through the GTFB module, so that information useful for face recognition can be fully extracted. The invention further discloses a low-resolution face recognition method and device. The low-resolution face recognition method comprises the steps of feature extraction, vector matching and the like, the accuracy of low-resolution face recognition is improved by efficiently extracting the face features, and the requirements of actual application are met.

Description

Technical field

[0001] The present invention belongs to the field of artificial intelligence, in particular, relating to a human face feature extraction method, low resolution face recognition method and equipment. Background technique

[0002] With the enhancement of computer power, deep learning technology is gradually applied to various fields of artificial intelligence, and has achieved breakthrough effects. In the face recognition direction, for high-resolution human face images, many neural network models (such as VGG, RESNet and Mobile "have been able to achieve very high recognition accuracy, but in actual monitoring scenarios, in order to get enough big enough Monitoring the field of view, the camera is usually installed high, and the face image resolution acquired by the camera is generally low, and the existing model is still low, and the accuracy of the low-resolution face image recognition is still low, and it needs to be further improved. Inventive content

[0003] ...

Examples

Embodiment 1

[0056] according to figure 1 The structure is built into the feature extracting network, wherein the initial feature extraction module 31 is 3 * 3 convolutional layer, the number of channels of the input and output initial feature extraction module 31 is 3 and 64, respectively, and the length and width of the image is unchanged. . The structure of the GTFB module 32 is like figure 2 As shown in the GTFB module 32, there is a residual connection, and the number of GTFB modules 32 is 6. In the GTFB module 32, the number of channels of input and output 3 * 3, 5 * 5 and the deformable convolutionary feature is equal to 64, 3 * 3 convolution and the deformable convolutionary output feature graph, and pass through 1 * 1 volume and activation function, the output channel is 64, the number of channels of the GTFB module 32 is 192, the number of channels of the volume is 192, the number of channels of the output is 64, such input and output the characteristic map of the GTFB module 32. The...

Embodiment 2

[0063] Based on the network in Example 1, the WSA Care Module 5 is added. For the internal portion of the WSA Care Module 5, the number of feature map channels output by 1 * 1 and the GTFB module 32 is first degraded to 1, and then the various characteristics obtained after the dimension are spliced, and then The number of feature drawings after 1 * 1 is reduced to 1, and after the SIGMOID function is activated, the space is given.

[0064]The comparative experiment is performed using the identical training and test conditions (including experimental methods and steps, hardware, frame, data set, optimizer, learning rate, etc.), and the results showed that the WSA attention module 5 was increased. For the 14 * 14 resolution FERET dataset, the correct rate of identification is 95.3%, and the correct rate of the recognition is 97.1% for the FERET dataset of 28 * 28 resolution. Note The WSA Care Module 5 has a good role in promoting the extraction of face characteristics and face reco...

Embodiment 3

[0066] On the basis of the network in Example 2, the hopping fusion module 6 is increased, and the feature extracts the network structure, such as Figure 4 Indicated. In the present embodiment, the hip fusion module 6 internal structure is Image 6 Indicated. After the first feature, the second feature inputs the jump fusion module 6, both first in the channel direction, obtain a first jump feature map of the number of channels 128. On the other hand, the first feature and the second feature are respectively passed through 3 * 3 convolution and the RELU activation function, respectively generate a second jump feature map and a third jump feature map of the channel number of 128, respectively. Then, the first jump feature, the second jump feature map, and the third hopping feature map are passed through element and fuse, and then pass through the 3 * 3 convolution and the RELU activation function, the output channel number is 128. . The final jump feature map and the feature map out...

Technical field

[0001] The present invention belongs to the field of artificial intelligence, in particular, relating to a human face feature extraction method, low resolution face recognition method and equipment. Background technique

[0002] With the enhancement of computer power, deep learning technology is gradually applied to various fields of artificial intelligence, and has achieved breakthrough effects. In the face recognition direction, for high-resolution human face images, many neural network models (such as VGG, RESNet and Mobile "have been able to achieve very high recognition accuracy, but in actual monitoring scenarios, in order to get enough big enough Monitoring the field of view, the camera is usually installed high, and the face image resolution acquired by the camera is generally low, and the existing model is still low, and the accuracy of the low-resolution face image recognition is still low, and it needs to be further improved. Inventive content

[0003] ...

Examples

Embodiment 1

[0056] according to figure 1 The structure is built into the feature extracting network, wherein the initial feature extraction module 31 is 3 * 3 convolutional layer, the number of channels of the input and output initial feature extraction module 31 is 3 and 64, respectively, and the length and width of the image is unchanged. . The structure of the GTFB module 32 is like figure 2 As shown in the GTFB module 32, there is a residual connection, and the number of GTFB modules 32 is 6. In the GTFB module 32, the number of channels of input and output 3 * 3, 5 * 5 and the deformable convolutionary feature is equal to 64, 3 * 3 convolution and the deformable convolutionary output feature graph, and pass through 1 * 1 volume and activation function, the output channel is 64, the number of channels of the GTFB module 32 is 192, the number of channels of the volume is 192, the number of channels of the output is 64, such input and output the characteristic map of the GTFB module 32. The...

Embodiment 2

[0063] Based on the network in Example 1, the WSA Care Module 5 is added. For the internal portion of the WSA Care Module 5, the number of feature map channels output by 1 * 1 and the GTFB module 32 is first degraded to 1, and then the various characteristics obtained after the dimension are spliced, and then The number of feature drawings after 1 * 1 is reduced to 1, and after the SIGMOID function is activated, the space is given.

[0064]The comparative experiment is performed using the identical training and test conditions (including experimental methods and steps, hardware, frame, data set, optimizer, learning rate, etc.), and the results showed that the WSA attention module 5 was increased. For the 14 * 14 resolution FERET dataset, the correct rate of identification is 95.3%, and the correct rate of the recognition is 97.1% for the FERET dataset of 28 * 28 resolution. Note The WSA Care Module 5 has a good role in promoting the extraction of face characteristics and face reco...

Embodiment 3

[0066] On the basis of the network in Example 2, the hopping fusion module 6 is increased, and the feature extracts the network structure, such as Figure 4 Indicated. In the present embodiment, the hip fusion module 6 internal structure is Image 6 Indicated. After the first feature, the second feature inputs the jump fusion module 6, both first in the channel direction, obtain a first jump feature map of the number of channels 128. On the other hand, the first feature and the second feature are respectively passed through 3 * 3 convolution and the RELU activation function, respectively generate a second jump feature map and a third jump feature map of the channel number of 128, respectively. Then, the first jump feature, the second jump feature map, and the third hopping feature map are passed through element and fuse, and then pass through the 3 * 3 convolution and the RELU activation function, the output channel number is 128. . The final jump feature map and the feature map out...

Images