Face image enhancement method and device for face scanning payment

Abstract

The invention discloses a face image enhancement method and device for face scanning payment, and relates to the technical field of image enhancement. The method comprises the following steps: acquiring a face image of a target low-light area; performing analysis according to the face image to obtain an image recognition coefficient; comparing and judging the image recognition coefficient with an image recognition threshold value; performing multi-angle fusion operation on the face image to obtain a first feature map; the first feature map is an image after multi-angle fusion operation, performing semantic segmentation operation on the first feature map to obtain a second feature map, and the second feature map is an image after semantic segmentation operation enhancement; according to the method, the face recognition accuracy and the overall image definition under the low-light environment condition are improved.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, specifically to a method and apparatus for enhancing facial images used in facial recognition payment. Background Technology

[0002] Image enhancement is an important branch of digital image processing, aiming to improve the visual quality of images or highlight specific features. With the advancement of computer technology and digital imaging, image enhancement has been widely used in fields such as medical imaging, remote sensing, security monitoring, and autonomous driving. Common enhancement methods include spatial domain processing (such as histogram equalization and filtering for noise reduction) and frequency domain processing (such as Fourier transform). In recent years, deep learning (such as generative adversarial networks) has significantly improved enhancement results, especially in low-light, blurred, or noisy scenes.

[0003] The prior art (CN118644885A) discloses a face recognition method and system based on neural network adaptive image enhancement, belonging to the field of image recognition technology. This method first evaluates the image quality of the original face image and determines the first face image based on the quality score. Then, it obtains the quality index of the first face image and determines an optimization strategy based on the quality index. Next, it adaptively enhances the first face image according to the optimization strategy to obtain an enhanced face image. Finally, it evaluates the similarity between the first face image and the enhanced face image based on contour features and local features. When the similarity meets the requirements, the enhanced face image is used for face recognition. This method solves the problem of face image optimization distortion, ensures the similarity of the face image before and after enhancement, and improves the accuracy of face recognition.

[0004] However, in low-light scenarios, the facial images are blurry and noisy, resulting in poor image quality and inaccurate feature extraction. This makes it impossible to accurately determine the image state, and the inaccurate feature extraction affects the accuracy and security of facial recognition payment. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of blurry and noisy facial images, which lead to poor image quality and inaccurate feature extraction, resulting in the inability to accurately determine the image state and affecting the accuracy and security of facial recognition payment. Therefore, this invention proposes a facial image enhancement method and device for facial recognition payment.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] First, a facial image enhancement method for facial recognition payment is proposed, the method comprising:

[0008] Acquire a face image of a target low-light region, wherein the face image is a face image of a low-light region;

[0009] Image recognition coefficients are obtained by analyzing facial images;

[0010] The image recognition coefficient is compared with the image recognition threshold for judgment;

[0011] A first feature map is obtained by performing a multi-angle fusion operation on the face image; the first feature map is the image after the multi-angle fusion operation; a second feature map is obtained by performing a semantic segmentation operation on the first feature map; the second feature map is the image after the semantic segmentation operation.

[0012] Optionally, image recognition coefficients are obtained by analyzing the face image, including:

[0013] The image data is divided into m blocks; the brightness, color, and noise of each block are extracted; and the brightness, color, and noise are normalized.

[0014] The specific method for obtaining image recognition coefficients is as follows:

[0015]

[0016] L j C1 represents the normalized brightness. j C2 represents the normalized color saturation. j N represents the normalized color contrast. j The noise is normalized, j = 1, 2, ..., m; m represents the total number of image blocks.

[0017] Optionally, the image recognition coefficients are compared with the image recognition threshold, including:

[0018] If the image recognition coefficient is less than the preset image recognition threshold, then image enhancement is required.

[0019] If the image recognition coefficient is greater than or equal to the preset image recognition threshold, then image enhancement is not required.

[0020] Optional, the multi-angle fusion process includes:

[0021] The face image is used as the input feature tensor; the input feature tensor is input into a deep convolutional layer for analysis to obtain a first feature tensor; the first feature tensor is sequentially input into a max pooling layer, a fully connected layer, and a sigmoid layer to obtain a second feature tensor; the first feature tensor is sequentially input into an average pooling layer, a fully connected layer, and a sigmoid layer to obtain a third feature tensor; the first feature tensor and the second feature tensor are fused to obtain a first fused tensor; the first feature tensor and the third feature tensor are fused to obtain a second fused tensor; the first feature tensor, the first fused tensor, and the second fused tensor are added to obtain a fourth feature tensor, which is denoted as the first feature map.

[0022] Optionally, the semantic segmentation operation includes:

[0023] The average brightness level B of category s is calculated using the formula. s and semantic smoothing loss YS;

[0024] Average brightness level B s How to obtain:

[0025]

[0026] α s Let α represent the set of pixels for category s, and n represent the set of pixels for category s. s Total number of pixels; B i l is the brightness level of the i-th pixel in the enhanced image; l represents the image input to the first feature map; s represents the semantic category of the image; S represents the total semantic category;

[0027] How to obtain the semantic smoothing loss YS:

[0028]

[0029] B s l It is the average brightness level of category s.

[0030] A facial image enhancement device for facial recognition payment is proposed, comprising:

[0031] Data acquisition module: Acquires facial images of a target low-light area, wherein the facial images are facial images of a low-light area;

[0032] Image recognition module: Analyzes facial images to obtain image recognition coefficients; compares the image recognition coefficients with the image recognition threshold for judgment;

[0033] Image enhancement module: Performs a multi-angle fusion operation on the face image to obtain a first feature map; the first feature map is the image after the multi-angle fusion operation; performs a semantic segmentation operation on the first feature map to obtain a second feature map; the second feature map is the image after the semantic segmentation operation.

[0034] Optionally, the image recognition module includes: an image segmentation module and a recognition analysis module.

[0035] The image segmentation module is used to divide the image data into m image blocks; extract the brightness, color, and noise of each image block; and normalize the brightness, color, and noise.

[0036] The recognition and analysis module is used to specifically obtain the image recognition coefficients.

[0037]

[0038] L j C1 represents the normalized brightness. j C2 represents the normalized color saturation. j N represents the normalized color contrast. j The noise is normalized, j = 1, 2, ..., m; m represents the total number of image blocks.

[0039] Optionally, the image recognition module includes: a comparison module:

[0040] The comparison module is used to perform image enhancement if the image recognition coefficient is less than a preset image recognition threshold, and not to perform image enhancement if the image recognition coefficient is greater than or equal to the preset image recognition threshold.

[0041] Optionally, the image enhancement module includes: a multi-angle data processing module and a fusion module.

[0042] The multi-angle data processing module is used to take the face image as an input feature tensor; input the input feature tensor into a deep convolutional layer for analysis to obtain a first feature tensor; input the first feature tensor into a max pooling layer, a fully connected layer and a sigmoid layer in sequence to obtain a second feature tensor; input the first feature tensor into an average pooling layer, a fully connected layer and a sigmoid layer in sequence to obtain a third feature tensor;

[0043] The fusion module is used to fuse the first feature tensor and the second feature tensor to obtain a first fusion tensor; fuse the first feature tensor and the third feature tensor to obtain a second fusion tensor; and add the first feature tensor, the first fusion tensor, and the second fusion tensor to obtain a fourth feature tensor, wherein the fourth feature tensor is denoted as the first feature map.

[0044] Optionally, the image enhancement module includes: an average brightness level module and a semantic smoothing loss module.

[0045] The average brightness level module is used to calculate the average brightness level B of category s using a formula. s and semantic smoothing loss YS;

[0046] Average brightness level B s How to obtain:

[0047]

[0048] α s Let α represent the set of pixels for category s, and n represent the set of pixels for category s. s Total number of pixels; B i l is the brightness level of the i-th pixel in the enhanced image; l represents the image input to the first feature map; s represents the semantic category of the image; S represents the total semantic category;

[0049] The semantic smoothing loss module is used to obtain the semantic smoothing loss YS.

[0050]

[0051] B s l It is the average brightness level of category s.

[0052] The beneficial effects of this invention are:

[0053] This invention proposes a facial image enhancement method for facial recognition payment. The method involves acquiring a facial image of a target low-light region; analyzing the facial image to obtain image recognition coefficients; comparing the image recognition coefficients with an image recognition threshold; performing a multi-angle fusion operation on the facial image to obtain a first feature map; and performing semantic segmentation on the first feature map to obtain a second feature map. By acquiring facial image data of a low-light region and analyzing the recognition coefficients, the current state of the image can be accurately determined. Furthermore, the multi-angle fusion operation enhances details and removes noise, improving the overall image clarity. By pushing the brightness of pixels of the same semantic category to their average brightness level, semantic smoothing loss improves the smoothness and consistency of each facial part. The combination of these two methods significantly improves the quality of the facial image and enhances the accuracy of feature extraction. Attached Figure Description

[0054] Figure 1 A flowchart of a facial image enhancement method for facial recognition payment provided in an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of a facial image enhancement device for facial recognition payment provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This invention provides a method for enhancing facial images for facial recognition payment. See also... Figure 1 , Figure 1 A flowchart illustrating a facial image enhancement method for facial recognition payment provided in an embodiment of the present invention. The method includes the following steps:

[0058] Acquire a face image of the target low-light region; the face image is a face image of the low-light region.

[0059] Image recognition coefficients are obtained by analyzing facial images;

[0060] The image recognition coefficient is compared with the image recognition threshold for judgment;

[0061] A first feature map is obtained by performing a multi-angle fusion operation on a face image; the first feature map is the image after the multi-angle fusion operation; a second feature map is obtained by performing a semantic segmentation operation on the first feature map; the second feature map is the image after the semantic segmentation operation.

[0062] The facial image enhancement method for facial recognition payment provided by this invention can accurately determine the current state of the image by acquiring facial image data in low-light areas and analyzing the recognition coefficients; furthermore, multi-angle fusion operations can enhance details and remove noise, improving the overall image clarity; by pushing the brightness of pixels of the same semantic category to their average brightness level, semantic smoothing loss improves the smoothness and consistency of each facial part; the combination of these two methods can significantly improve the quality of facial images and enhance the accuracy of feature extraction.

[0063] In one implementation, image recognition coefficients are obtained by analyzing facial image data, including:

[0064] The face image data is divided into m blocks; the brightness, color, and noise of each block are extracted; and the brightness, color, and noise are normalized.

[0065] The specific method for obtaining image recognition coefficients is as follows:

[0066]

[0067] Lj C1 represents the normalized brightness. j C2 represents the normalized color saturation. j N represents the normalized color contrast. j The noise is normalized, j = 1, 2, ..., m; m represents the total number of image blocks.

[0068] Specifically, it should be noted that the normalization method is Z-Score normalization, etc.; the image recognition coefficient represents a numerical indicator that quantifies the image quality; the higher the brightness, the clearer the image, the more effective information it provides, and the higher the corresponding image recognition coefficient; the higher the color saturation, the more vivid the color, the greater the contrast, the more prominent the image color features, and the higher the corresponding image recognition coefficient; the larger the noise Nj, the worse the image quality, the more interference information, and the lower the corresponding image recognition coefficient.

[0069] In one implementation, by segmenting and extracting multi-dimensional features from facial image data, the image can be analyzed more meticulously and comprehensively, making image recognition more accurate; brightness, color, and noise are normalized; and image recognition coefficients are calculated using specific formulas to provide quantitative basis for image recognition, thereby enhancing the stability and adaptability of the recognition system.

[0070] In one implementation, the image recognition coefficient is compared with the image recognition threshold, including:

[0071] If the image recognition coefficient is less than the preset image recognition threshold, then image enhancement is required.

[0072] If the image recognition coefficient is greater than or equal to the preset image recognition threshold, then image enhancement is not required.

[0073] Specifically, it should be noted that the preset image recognition threshold is obtained by staff based on historical experience. By comparing the image recognition coefficient with the preset threshold, the system intelligently determines whether image enhancement is needed. This automatic decision-making process eliminates the need for manual intervention, improves processing efficiency, ensures image quality meets standards, enhances subsequent recognition accuracy, and strengthens the system's overall intelligence.

[0074] In one implementation, the multi-angle fusion operation process includes:

[0075] The face image is used as the input feature tensor; the input feature tensor is fed into a deep convolutional layer for analysis to obtain the first feature tensor; the first feature tensor is sequentially fed into a max pooling layer, a fully connected layer, and a sigmoid layer to obtain the second feature tensor; the first feature tensor is sequentially fed into an average pooling layer, a fully connected layer, and a sigmoid layer to obtain the third feature tensor; the first feature tensor and the second feature tensor are fused to obtain the first fused tensor; the first feature tensor and the third feature tensor are fused to obtain the second fused tensor; the first feature tensor, the first fused tensor, and the second fused tensor are added to obtain the fourth feature tensor, which is denoted as the first feature map.

[0076] Specifically, it should be noted that the first feature tensor is obtained by processing through a deep convolutional layer; the second feature tensor is obtained by processing through a max pooling layer, a fully connected layer, and a sigmoid layer; and the third feature tensor is obtained by processing through an average pooling layer, a fully connected layer, and a sigmoid layer.

[0077] In one implementation method, the multi-angle fusion operation specifically includes:

[0078] Max pooling layers are used to capture prominent features in the image; the most prominent features include facial features, face contour features, etc. The prominent structure is obtained using the formula: M = MaxPool(Dconv(F)); where M represents the prominent structure, Dconv(F) = P(D(F)) is a depthwise separable convolution, D represents a depthwise convolution, and F is the input feature tensor. Adaptive coefficients S are generated through fully connected layers and sigmoid layers. max =sigmoid(FC(M)), where FC represents a fully connected layer; finally, the salient features are obtained as follows: Y max =S max ×F;

[0079] Average pooling is used to smooth the feature distribution and suppress noise; A = AvgPool(Dconv(F)); A is the average feature; similarly, the average feature generates adaptive coefficients: S avg = sigmoid(FC(A)), then the balance characteristic is: Y avg =S avg ×F;

[0080] The calculation formula for multi-angle feature fusion is: Y = Y max +Y avg +F.

[0081] In one implementation, multi-view fusion operations ensure that the enhancement goes beyond simple brightness adjustment, improving overall image quality by capturing fine details while suppressing noise. Therefore, this significantly enhances the network's ability to process low-light images and improves the performance of low-light image enhancement.

[0082] In one implementation, a semantic segmentation operation is performed on the first feature map to obtain a second feature map, including:

[0083] The average brightness level B of category s is calculated using the formula. s and semantic smoothing loss YS;

[0084] Average brightness level B s How to obtain:

[0085]

[0086] α s Let α represent the set of pixels for category s, and n represent the set of pixels for category s. s Total number of pixels; B i l is the brightness level of the i-th pixel in the enhanced image; l represents the image input to the first feature map; s represents the semantic category of the image; S represents the total semantic category;

[0087] How to obtain the semantic smoothing loss YS:

[0088]

[0089] B s l It is the average brightness level of category s.

[0090] In one implementation, for example in low-light enhancement, the semantic segmentation operation assigns a category label to each pixel to ensure a smooth transition in brightness for pixels of the same category (such as skin, eyes, etc.) and avoid localized over-darkness.

[0091] In one implementation, semantic smoothing loss improves the smoothness and consistency of each facial part by pushing the pixel brightness of the same semantic category to its average brightness level; it effectively avoids the problem of uneven local exposure, making the enhanced image smoother and more consistent within the semantic category.

[0092] Based on the same inventive concept, this invention also provides a facial image enhancement device for facial recognition payment. See also Figure 2 , Figure 2 A schematic diagram of a facial image enhancement device for facial recognition payment provided in an embodiment of the present invention includes:

[0093] Data acquisition module: Acquires facial images of the target low-light area; the facial images are facial images of the low-light area.

[0094] Image recognition module: Analyzes facial images to obtain image recognition coefficients; compares the image recognition coefficients with the image recognition threshold for judgment;

[0095] Image enhancement module: Performs multi-angle fusion operation on face image to obtain first feature map; first feature map is image after multi-angle fusion operation; performs semantic segmentation operation on first feature map to obtain second feature map; second feature map is image after semantic segmentation operation enhancement.

[0096] The facial image enhancement device for facial recognition payment provided by this invention can accurately determine the current state of the image by acquiring facial image data in low-light areas and analyzing the recognition coefficients; furthermore, it can enhance details and remove noise through multi-angle fusion operations, thereby improving the overall clarity of the image; by pushing the brightness of pixels of the same semantic category to their average brightness level, semantic smoothing loss improves the smoothness and consistency of each facial part; the combination of the two can significantly improve the quality of facial images and enhance the accuracy of feature extraction.

[0097] It should be noted that, in this document, terms such as “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for enhancing facial images for facial recognition payment, characterized in that, The method includes: Acquire a face image of a target low-light region, wherein the face image is a face image of a low-light region; Image recognition coefficients are obtained by analyzing facial images; The image recognition coefficient is compared with the image recognition threshold for judgment; A first feature map is obtained by performing a multi-angle fusion operation on the face image; the first feature map is the image after the multi-angle fusion operation; a second feature map is obtained by performing a semantic segmentation operation on the first feature map; the second feature map is the image after the semantic segmentation operation.

2. The facial image enhancement method for facial recognition payment according to claim 1, characterized in that, The process of analyzing facial images to obtain image recognition coefficients includes: The image data is divided into m blocks; the brightness, color, and noise of each block are extracted; and the brightness, color, and noise are normalized. The specific method for obtaining image recognition coefficients is as follows: L j C1 represents the normalized brightness. j C2 represents the normalized color saturation. j N represents the normalized color contrast. j The noise is normalized, j = 1, 2, ..., m; m represents the total number of image blocks.

3. The facial image enhancement method for facial recognition payment according to claim 1, characterized in that, The step of comparing and judging the image recognition coefficient with the image recognition threshold includes: If the image recognition coefficient is less than the preset image recognition threshold, then image enhancement is required. If the image recognition coefficient is greater than or equal to the preset image recognition threshold, then image enhancement is not required.

4. The facial image enhancement method for facial recognition payment according to claim 1, characterized in that, The process of the multi-angle fusion operation includes: The face image is used as the input feature tensor; the input feature tensor is input into a deep convolutional layer for analysis to obtain a first feature tensor; the first feature tensor is sequentially input into a max pooling layer, a fully connected layer, and a sigmoid layer to obtain a second feature tensor; the first feature tensor is sequentially input into an average pooling layer, a fully connected layer, and a sigmoid layer to obtain a third feature tensor; the first feature tensor and the second feature tensor are fused to obtain a first fused tensor; the first feature tensor and the third feature tensor are fused to obtain a second fused tensor; the first feature tensor, the first fused tensor, and the second fused tensor are added to obtain a fourth feature tensor, which is denoted as the first feature map.

5. A facial image enhancement method for facial recognition payment according to claim 1, characterized in that, The semantic segmentation operation includes the following steps: The average brightness level B of category s is calculated using the formula. s and semantic smoothing loss YS; Average brightness level B s How to obtain: α s Let α represent the set of pixels for category s, and n represent the set of pixels for category s. s Total number of pixels; B i l is the brightness level of the i-th pixel in the enhanced image; l represents the image input to the first feature map; s represents the semantic category of the image; S represents the total semantic category; How to obtain the semantic smoothing loss YS: B s l It is the average brightness level of category s.

6. A facial image enhancement device for facial recognition payment, characterized in that, The device includes: Data acquisition module: Acquires facial images of a target low-light area, wherein the facial images are facial images of a low-light area; Image recognition module: Analyzes facial images to obtain image recognition coefficients; compares the image recognition coefficients with the image recognition threshold for judgment; Image enhancement module: Performs a multi-angle fusion operation on the face image to obtain a first feature map; the first feature map is the image after the multi-angle fusion operation; performs a semantic segmentation operation on the first feature map to obtain a second feature map; the second feature map is the image after the semantic segmentation operation.

7. A facial image enhancement device for facial recognition payment according to claim 6, characterized in that, The image recognition module includes: an image segmentation module and a recognition analysis module. The image segmentation module is used to divide the image data into m image blocks; extract the brightness, color, and noise of each image block; and normalize the brightness, color, and noise. The recognition and analysis module is used to specifically obtain the image recognition coefficients. L j C1 represents the normalized brightness. j C2 represents the normalized color saturation. j N represents the normalized color contrast. j The noise is normalized, j = 1, 2, ..., m; m represents the total number of image blocks.

8. A facial image enhancement device for facial recognition payment according to claim 6, characterized in that, The image recognition module includes: a comparison module. The comparison module is used to perform image enhancement if the image recognition coefficient is less than a preset image recognition threshold, and not to perform image enhancement if the image recognition coefficient is greater than or equal to the preset image recognition threshold.

9. A facial image enhancement device for facial recognition payment according to claim 6, characterized in that, The image enhancement module includes: a multi-angle data processing module and a fusion module. The multi-angle data processing module is used to take the face image as an input feature tensor; input the input feature tensor into a deep convolutional layer for analysis to obtain a first feature tensor; input the first feature tensor into a max pooling layer, a fully connected layer and a sigmoid layer in sequence to obtain a second feature tensor; input the first feature tensor into an average pooling layer, a fully connected layer and a sigmoid layer in sequence to obtain a third feature tensor; The fusion module is used to fuse the first feature tensor and the second feature tensor to obtain a first fusion tensor; fuse the first feature tensor and the third feature tensor to obtain a second fusion tensor; and add the first feature tensor, the first fusion tensor, and the second fusion tensor to obtain a fourth feature tensor, wherein the fourth feature tensor is denoted as the first feature map.

10. A facial image enhancement device for facial recognition payment according to claim 6, characterized in that, The image enhancement module includes: an average brightness level module and a semantic smoothing loss module. The average brightness level module is used to calculate the average brightness level B of category s using a formula. s and semantic smoothing loss YS; Average brightness level B s How to obtain: α s Let α represent the set of pixels for category s, and n represent the set of pixels for category s. s Total number of pixels; B i l is the brightness level of the i-th pixel in the enhanced image; l represents the image input to the first feature map; s represents the semantic category of the image; S represents the total semantic category; The semantic smoothing loss module is used to obtain the semantic smoothing loss YS. B s l It is the average brightness level of category s.

Images