Iris recognition and face image enhancement integrated processing method

Abstract

This invention relates to the field of digital image processing technology, providing an integrated processing method for iris recognition and face image enhancement, comprising the following steps: S1, acquiring the original image to be processed containing the face region, and performing size normalization and grayscale preprocessing on the original image to obtain a standardized input image; S2, inputting the standardized input image into a shared feature encoding unit, performing multi-scale feature extraction, and outputting a shared encoded feature map containing global facial structural features and local iris texture features; S3, inputting the shared encoded feature map into the face image enhancement branch and the iris recognition branch respectively, and performing parallel collaborative decoding. By adopting a shared encoding and dual-branch collaborative architecture, face enhancement and iris recognition are integrated, and the lightweight design is suitable for real-time deployment at the edge; at the same time, iris texture is preserved with dual-constraint loss, combined with quality-aware dynamic fusion features, balancing recognition accuracy and robustness, and providing greater stability in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, specifically to an integrated processing method for iris recognition and face image enhancement. Background Technology

[0002] With the widespread adoption of identity authentication in scenarios such as access control, security, smart terminals, and examination room verification, multi-biometric feature fusion recognition has become the mainstream technology due to its security and robustness advantages. Face and iris, as core features for non-contact data acquisition, face and iris recognition face higher demands on image quality and recognition accuracy in real-world scenarios with complex lighting, pose shifts, and partial occlusion. Integrated collaborative processing has become an industry trend.

[0003] Currently, face enhancement and iris recognition are mostly independent modules with non-shared feature encoding and redundant computational processes, resulting in large model parameters and difficulty in real-time operation at edge devices. Furthermore, image enhancement easily damages iris texture, and feature fusion uses fixed weights, which cannot adapt to fluctuations in image quality. Ultimately, this leads to problems such as low recognition accuracy in low-quality scenes, deployment lag at edge devices, and insufficient anti-interference capabilities, making it difficult to meet practical application requirements. To address these technical issues, this application proposes an integrated processing method for iris recognition and face image enhancement. Summary of the Invention To address the shortcomings of existing technologies, this invention provides an integrated processing method for iris recognition and face image enhancement, which solves the problems of independent separation, computational redundancy, easy damage to iris texture, and poor recognition robustness in existing technologies.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an integrated iris recognition and face image enhancement processing method, comprising the following steps: S1. Acquire the original image to be processed containing the face region, and perform size normalization and grayscale preprocessing on the original image to be processed to obtain a standardized input image. S2. Input the standardized input image into the shared feature coding unit, perform multi-scale feature extraction, and output a shared coded feature map containing global facial structural features and local iris texture features; S3. Input the shared coded feature map into the face image enhancement branch and the iris recognition branch respectively, and perform parallel collaborative decoding. The face image enhancement branch completes adaptive enhancement of the face image based on the shared coded feature map and outputs the enhanced face image. The iris recognition branch completes iris region localization and iris feature extraction simultaneously based on the shared coded feature map and outputs the standard iris feature vector. S4. Calculate the image quality evaluation factor of the original image to be processed. Based on the image quality evaluation factor, adaptively adjust the fusion weight of the global facial features and the standard iris feature vector corresponding to the enhanced face image to generate a fused identity feature vector. Match the fused identity feature vector with the pre-stored registration feature library and output the identity recognition result.

[0005] Preferably, the size normalization process in S1 specifically involves scaling the original image to be processed to a resolution of 640×480 pixels. The grayscale conversion uses a weighted average method, and the grayscale conversion formula is as follows: in, This represents the grayscale value of the pixel after grayscale conversion. , , These are the red, green, and blue channel component values ​​of the corresponding pixels in the original color image.

[0006] Preferably, the shared feature encoding unit in step S2 adopts a lightweight MobileNetV2backbone architecture, which includes 4 downsampling modules and 8 inverse residual bottleneck modules; wherein, the first 2 downsampling modules extract global facial structural features, and the last 2 downsampling modules and 8 inverse residual bottleneck modules extract local iris texture features, and the output shared encoded feature map has 128 channels and a feature map size of 40×30 pixels.

[0007] Preferably, the face image enhancement branch in S3 includes a multi-scale feature fusion module and an upsampling decoding module. The optimization objective of the face image enhancement branch is a double-constraint loss function, with the following formula: in, For the total loss function, For pixel-level reconstruction loss, Preserve loss for iris texture , To balance the weighting coefficients, and , The value range is 0.4 to 0.6. The value range is 0.4 to 0.6.

[0008] Preferably, the iris texture retention loss The calculation formula is: in, This is the function for calculating the mean square error. To enhance the feature vector of the iris region in the post-face image, This is the baseline feature vector corresponding to a standard clear iris image.

[0009] Preferably, the iris recognition branch in S3 sequentially includes an iris region localization module, an iris normalization module, and an iris feature extraction module. The iris region localization module uses the Hough circle transform algorithm based on a shared coded feature map to accurately locate the inner and outer boundaries of the iris, with the localization deviation threshold controlled within 2 pixels. The iris normalization module expands the located annular iris region into a rectangular normalized iris image with a size of 512×64 pixels. The iris feature extraction module extracts a standard iris feature vector with a dimension of 256 based on the normalized iris image.

[0010] Preferably, the image quality evaluation factor in S4 The calculation formula is: in, To score the uniformity of illumination, Rate the image sharpness. Scoring facial pose deviation. Score the degree of face occlusion. , , , For the corresponding weight coefficients, and , The range of values ​​is , A higher value indicates better image quality.

[0011] Preferably, in step S4, the image quality evaluation factor is used. The specific method for adaptively adjusting the fusion weights is as follows: Fusion weights of global facial features Fusion weights of iris feature vectors Fusion of identity feature vectors The calculation formula is: in, To enhance the 256-dimensional global feature vector of the face extracted from the post-face image, It is a 256-dimensional standard iris feature vector.

[0012] Preferably, the feature matching in S4 specifically involves: calculating the cosine similarity between the fused identity feature vector and each registered feature vector in the registration feature library; when the maximum cosine similarity is greater than the preset similarity threshold of 0.85, the identity recognition result of successful matching is output; otherwise, the recognition result of failed matching is output.

[0013] The integrated iris recognition and face image enhancement system includes: an image preprocessing unit, which acquires the original image to be processed, performs size normalization and grayscale preprocessing, and outputs a standardized input image; A shared feature coding unit is communicatively connected to the image preprocessing unit and is used to perform multi-scale feature extraction on the standardized input image and output a shared coded feature map. The dual-branch collaborative decoding unit, which is communicatively connected to the shared feature encoding unit, includes a face image enhancement branch module and an iris recognition branch module, which are used to perform adaptive face image enhancement and iris feature extraction in parallel, and output the enhanced face image and standard iris feature vector respectively. The quality-perceived fusion and recognition matching unit is communicatively connected to the dual-branch collaborative decoding unit and the image preprocessing unit, respectively. It is used to calculate the image quality evaluation factor, adaptively adjust the feature fusion weights and generate a fused identity feature vector, and complete the output of feature matching and identity recognition results.

[0014] Preferably, the system is deployed on an edge intelligent device, the computing power of the edge intelligent device is ≥1 TOPS, the system single-frame image processing latency is ≤33ms, and the real-time processing frame rate is ≥30fps.

[0015] This invention provides an integrated method for iris recognition and face image enhancement. It offers the following advantages: 1. This invention achieves process reuse and parameter sharing for face image enhancement and iris recognition by adopting an integrated architecture of shared feature encoding and dual-branch collaborative decoding, significantly reducing the number of model parameters and computational overhead. Multi-scale feature extraction is completed through a lightweight MobileNetV2 backbone, with a single feature map simultaneously carrying both global facial structure and local iris texture features. Combined with a parallel decoding architecture, real-time processing with a single-frame processing latency of ≤33ms and a frame rate of ≥30fps can be achieved on edge devices with a computing power ≥1 TOPS. Further grayscale optimization in the preprocessing stage further reduces computational load, adapting to the lightweight deployment and real-time operation requirements of edge devices.

[0016] 2. This invention employs a dual-constraint loss function to fully preserve the core texture features of the iris while adaptively enhancing the face image, thus avoiding negative impacts on recognition accuracy caused by the enhancement operation. Simultaneously, based on multi-dimensional image quality evaluation factors, it achieves adaptive dynamic weighted fusion of global facial features and iris features. The weights can be dynamically adjusted according to image conditions such as illumination, sharpness, pose, and occlusion. This leverages the high accuracy and strong anti-spoofing advantages of iris recognition while supplementing robustness in extreme scenarios through facial features, achieving a dual improvement in accuracy and anti-interference capability for identity recognition in complex environments. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] 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.

[0019] Example: Please see the appendix Figure 1 This invention provides an integrated iris recognition and face image enhancement processing method, comprising the following steps: S1. Acquire the original image to be processed containing the face region, and perform size normalization and grayscale preprocessing on the original image to be processed to obtain a standardized input image. The size normalization process in S1 is as follows: the original image to be processed is uniformly scaled to a resolution of 640×480 pixels. The grayscale conversion uses a weighted average method, and the grayscale conversion formula is as follows: in, This represents the grayscale value of the pixel after grayscale conversion. , , These are the red, green, and blue channel component values ​​of the corresponding pixels in the original color image; Specifically, the raw image is acquired using a CMOS visible light image sensor built into the edge-end intelligent device, with a frame rate set to 30fps, compatible with mainstream resolutions such as 1080P and 720P. After acquisition, the effective facial region in the image is quickly located using the MTCNN face detection algorithm. The region containing the complete face is cropped, and meaningless background areas are removed. Then, a bicubic interpolation algorithm is used to uniformly scale the cropped face image to a fixed resolution of 640×480 pixels, completing the size normalization process and eliminating size differences in the input image. Subsequently, a weighted average grayscale processing is performed on the normalized color image. The grayscale value is calculated pixel by pixel according to the above grayscale conversion formula, converting the three-channel color image into a single-channel grayscale image, eliminating redundant information in the color channels, and obtaining a standardized input image. While retaining the core features of the face and iris, the computational load of the subsequent feature extraction stage is reduced to 1 / 3 of the original three-channel processing mode.

[0020] Please see the appendix Figure 1 S2. The standardized input image is input into the shared feature coding unit for multi-scale feature extraction, and the output is a shared coded feature map containing global facial structural features and local iris texture features. The shared feature coding unit in step S2 adopts a lightweight MobileNetV2backbone architecture, which includes 4 downsampling modules and 8 inverse residual bottleneck modules. The first 2 downsampling modules extract global facial structural features, and the last 2 downsampling modules and 8 inverse residual bottleneck modules extract local iris texture features. The output shared coded feature map has 128 channels and a feature map size of 40×30 pixels. Specifically, after the standardized input image is input into the shared feature encoding unit, it first enters the first 3×3 convolutional layer for initial feature mapping, with 32 output channels and a feature map size adjusted to 320×240 pixels, completing the initial feature extraction. Subsequently, it passes through four downsampling modules with a stride of 2, where the first and second downsampling modules perform 2x and 4x downsampling respectively, outputting feature map sizes of 160×120 pixels and 80×60 pixels. This stage employs a depthwise separable convolutional structure, focusing on extracting facial contours, spatial distribution of facial features, and surface features. The first step is to extract global facial structural features, such as the overall contour, to fully preserve the overall spatial semantic information of the face. The third and fourth downsampling modules further perform 8x and 16x downsampling, respectively, resulting in a final output feature map with a basic size of 40×30 pixels. This is followed by eight cascaded inverse residual bottleneck modules. Each bottleneck module adopts a core structure of "1×1 pointwise convolution for dimensionality upscaling - 3×3 depthwise convolution - 1×1 pointwise convolution for dimensionality downscaling," with an expansion factor of 6. This achieves accurate extraction of subtle local features such as iris texture details, pupil boundaries, and iris ring textures with extremely low computational cost. Finally, a channel fusion layer is used to stitch and fuse multi-scale features, outputting a shared encoded feature map with 128 channels and a size of 40×30 pixels. This allows a single feature map to simultaneously carry dual feature information of the global facial structure and local iris texture, enabling model parameter sharing and computational reuse in the encoding stage, significantly reducing the overall algorithm's parameter and computational cost.

[0021] Please see the appendix Figure 1 S3. The shared coded feature map is input into the face image enhancement branch and the iris recognition branch respectively, and parallel collaborative decoding is performed. The face image enhancement branch performs adaptive enhancement of the face image based on the shared coded feature map and outputs the enhanced face image. The iris recognition branch simultaneously performs iris region localization and iris feature extraction based on the shared coded feature map and outputs a standard iris feature vector. The face image enhancement branch in S3 includes a multi-scale feature fusion module and an upsampling decoding module. The optimization objective of the face image enhancement branch is a double-constraint loss function, the formula of which is: in, For the total loss function, For pixel-level reconstruction loss, Preserve loss for iris texture , To balance the weighting coefficients, and , The value range is 0.4 to 0.6. The value ranges from 0.4 to 0.6, representing the loss in iris texture preservation. The calculation formula is: in, This is the function for calculating the mean square error. To enhance the feature vector of the iris region in the post-face image, To correspond to the baseline feature vector of a standard clear iris image, the iris recognition branch in S3 includes an iris region localization module, an iris normalization module, and an iris feature extraction module. The iris region localization module uses the Hough circle transform algorithm based on the shared coded feature map to accurately locate the inner and outer boundaries of the iris, with the localization deviation threshold controlled within 2 pixels. The iris normalization module expands the located annular iris region into a rectangular normalized iris image with a size of 512×64 pixels. The iris feature extraction module extracts a standard iris feature vector with a dimension of 256 based on the normalized iris image. Specifically, the shared encoded feature map is synchronously input into the face image enhancement branch and the iris recognition branch. The two branches adopt a parallel computing architecture and achieve synchronous decoding processing on the edge device. The decoding latency of a single branch does not exceed 15ms, ensuring the real-time performance of the overall processing flow. For the face image enhancement branch, firstly, a multi-scale feature fusion module is used to perform skip connections and channel stitching between the shared encoded feature map and the feature maps of different levels output by the four downsampling modules in the encoding stage, fusing multi-scale feature information from different receptive fields to compensate for details and edge information lost during downsampling. Subsequently, through four sets of cascaded upsampling decoding modules, each of which contains a 2x bilinear interpolation upsampling layer and a 3×3 convolutional layer, the feature map size is gradually restored from 40×30 pixels to 640×480 pixels. Finally, a single-channel enhanced face image with a pixel value range of 0-255 is output through a Sigmoid activation layer, achieving adaptive brightness correction, detail enhancement, and noise suppression for low-quality face images such as those in low light, blurred, or backlit conditions. During the model training phase, this branch uses the aforementioned double-constraint loss function as the optimization objective. In this embodiment, the balancing weight coefficients α and β are set to 0.5, where the pixel-level reconstruction loss... Mean squared error loss is used to calculate the pixel-by-pixel grayscale difference between the enhanced image and the standard clear face image, constraining the reconstruction effect of the overall face structure, brightness distribution, and contour; iris texture preservation loss is also employed. A pre-trained lightweight iris feature extraction network is used to extract 256-dimensional feature vectors of the iris region from both enhanced and standard clear images. The mean square error of the two feature vectors is calculated to enforce that the core texture features of the iris are not destroyed during image enhancement, thus avoiding negative impacts on the accuracy of subsequent iris recognition. For the iris recognition branch, the iris region localization module first performs adaptive threshold binarization and Canny edge detection on the shared encoded feature map to separate the pupil, iris, and sclera regions. Then, based on the Hough circle transform algorithm, the inner boundary circular contour of the pupil and the outer boundary circular contour of the iris are fitted respectively, while removing interference areas such as eyelids, eyelashes, and light spots. The localization deviation threshold of the inner and outer boundaries is strictly controlled within 2 pixels to achieve accurate segmentation of the effective iris region. Subsequently, the iris normalization module uses a polar coordinate-Cartesian coordinate transformation algorithm, with the pupil center as the origin, to normalize the circular iris region. The domain performs 360° polar angle sampling and radial sampling to expand the annular iris region into a rectangular normalized iris image with a fixed size of 512×64 pixels, completely eliminating the differences in iris size and angle caused by shooting distance, pupil dilation, and face rotation. Finally, the iris feature extraction module performs texture filtering on the normalized iris image at 4 scales and 8 directions based on the Log-Gabor filter to extract the local phase feature information of the iris. Then, through feature encoding and PCA dimensionality reduction processing, the output is a standard iris feature vector with a fixed dimension of 256, which is used for subsequent identity feature fusion and matching.

[0022] Please see the appendix Figure 1 S4. Calculate the image quality evaluation factor of the original image to be processed. Based on the image quality evaluation factor, adaptively adjust the fusion weights of the global facial features and the standard iris feature vector corresponding to the enhanced face image to generate a fused identity feature vector. Match the fused identity feature vector with the pre-stored registration feature library to output the identity recognition result. The image quality evaluation factor in S4... The calculation formula is: in, To score the uniformity of illumination, Rate the image sharpness. Scoring facial pose deviation. Score the degree of face occlusion. , , , For the corresponding weight coefficients, and , The range of values ​​is , A higher value indicates better image quality. S4 is based on image quality evaluation factors. The specific method for adaptively adjusting the fusion weights is as follows: Fusion weights of global facial features Fusion weights of iris feature vectors Fusion of identity feature vectors The calculation formula is: in, To enhance the 256-dimensional global feature vector of the face extracted from the post-face image, Given a 256-dimensional standard iris feature vector, feature matching in S4 specifically involves calculating the cosine similarity between the fused identity feature vector and each registered feature vector in the registration feature library. When the maximum cosine similarity is greater than the preset similarity threshold of 0.85, the identity recognition result of a successful match is output; otherwise, the recognition result of a failed match is output. Specifically, firstly, for the original image to be processed, the image quality evaluation factor Q is calculated, where the weight coefficients of each component are set as follows: Take 0.25, Take 0.3, Take 0.25, A score of 0.2 was used, and all sub-scores were normalized to the [0,1] interval. Among these, the illumination uniformity score... Image sharpness score is obtained by calculating the global variance of the image's grayscale values. A smaller variance indicates a more uniform illumination distribution, corresponding to a higher score. The sum of gradient magnitudes of an image is calculated using the Laplacian gradient operator. A larger gradient value indicates sharper image edges and richer details, resulting in a higher score. Face pose deflection scoring is also included. A 68-point facial landmark detection algorithm is used to locate key facial feature points and calculate the tilt, yaw, and roll angles. Smaller tilt angles result in higher scores; when the tilt angle exceeds 30° in any direction, the score is set to 0. (Face occlusion score is also included.) The score is calculated by detecting the pixel occlusion ratio of the core areas of the face, including the eyes, mouth, and nose. A lower occlusion ratio results in a higher score; when the iris occlusion ratio exceeds 50%, the score is set to 0. Based on the above parameters and formulas, the image quality evaluation factor Q is calculated, enabling a comprehensive quantitative evaluation of the original input image quality. Subsequently, the feature fusion weights are adaptively adjusted based on the Q value to achieve dynamic weighted fusion: for example, when the original image has sufficient lighting, no occlusion, and high clarity, and the calculated Q value is 0.9, the global facial feature fusion weights are adjusted accordingly. The weight for iris feature fusion is 0.1. The Q-value is 0.9, prioritizing iris features for identity authentication due to their higher recognition accuracy and stronger anti-spoofing capabilities. However, when the original image experiences severe backlighting or the iris area is obscured by sunglasses, resulting in a Q-value of 0.2, the global facial feature fusion weights are adjusted accordingly. The weight for iris feature fusion is 0.8. The weighting is set to 0.2, significantly increasing the proportion of facial features and avoiding recognition errors caused by low-quality iris features, thus adapting to recognition needs in complex scenarios. Simultaneously, for the enhanced facial image output from step S3, a feature extraction network derived from the same source as the shared feature encoding unit is used to extract a 256-dimensional global facial feature vector. , compared with the 256-dimensional standard iris feature vector output by the iris recognition branch The weighted fusion is performed according to the above weights, generating a fused identity feature vector with a dimension of 256. To ensure consistency in feature dimensions while balancing computational efficiency and robustness, a feature matching operation is performed. The pre-stored registration feature library contains 256-dimensional standard fused identity feature vectors of registered users. The cosine similarity between the fused identity feature vector to be matched and each registered feature vector in the library is calculated. The cosine similarity value ranges from [0,1], with higher values ​​indicating a higher degree of matching. When the calculated maximum cosine similarity is greater than a preset threshold of 0.85, the feature match is considered successful, and the user's ID and related information corresponding to that registration feature are output. When the calculated maximum cosine similarity is less than or equal to 0.85, the feature match is considered unsuccessful, and the recognition result "Identity Unauthorized" is output. This completes the integrated processing and identity recognition of a single frame image.

[0023] The integrated iris recognition and face image enhancement system includes: an image preprocessing unit, which acquires the original image to be processed, performs size normalization and grayscale preprocessing, and outputs a standardized input image; The shared feature coding unit is connected in communication with the image preprocessing unit and is used to perform multi-scale feature extraction on the standardized input image and output a shared coded feature map. The dual-branch collaborative decoding unit is connected to the shared feature encoding unit and includes a face image enhancement branch module and an iris recognition branch module. It is used to complete the adaptive enhancement of face images and iris feature extraction in parallel, and output the enhanced face image and standard iris feature vector respectively. The quality perception fusion and recognition matching unit is connected to the dual-branch collaborative decoding unit and the image preprocessing unit respectively. It is used to calculate the image quality evaluation factor, adaptively adjust the feature fusion weight and generate the fused identity feature vector, and complete the output of feature matching and identity recognition results. The system is deployed on edge intelligent devices with a computing power of ≥1TOPS, a single frame image processing latency of ≤33ms, and a real-time processing frame rate of ≥30fps. Specifically, a modular hardware and software co-design architecture is adopted. At the hardware level, it is deployed on edge smart devices equipped with ARM Cortex-A76 architecture processors. The devices integrate a 2-megapixel CMOS visible light image sensor and a built-in NPU neural network computing unit, with a total computing power of 2 TOPS, meeting the minimum operating requirement of ≥1 TOPS computing power for the system. At the software level, it is developed based on an embedded Linux system and uses the TensorFlow Lite lightweight inference framework to perform INT8 quantization compression on the algorithm models of each unit in the above method, compressing the overall model size to less than 8MB. Real-time inference is achieved at the edge without relying on cloud server computing, which not only ensures the privacy and security of facial and iris biometric data, but also significantly improves the real-time performance of the processing flow.

[0024] The image preprocessing unit is directly connected to the device's built-in image sensor, completing real-time acquisition of raw images, face region detection, size normalization, and grayscale processing, providing standardized input for subsequent units. The shared feature encoding unit, dual-branch collaborative decoding unit, and quality-aware fusion and recognition matching unit are all deployed on the device's NPU computing unit. Through pipelined computing scheduling, the entire process latency for a single 640×480 pixel image is ≤30ms, meeting the performance requirement of a real-time processing frame rate ≥30fps. It can be directly adapted to various real-time identity recognition scenarios such as access control gates, smart locks, mobile identity authentication, and examination room identity verification. Simultaneously, the system reserves standard RS485, Ethernet, and Wiegand data interfaces, allowing direct connection to third-party business systems such as access control, attendance management, and security alerts, synchronizing identity recognition results to these systems in real time, achieving integrated identity authentication and business linkage.

[0025] 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, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An integrated processing method of iris recognition and face image enhancement, characterized in that, Includes the following steps: S1. Acquire the original image to be processed containing the face region, and perform size normalization and grayscale preprocessing on the original image to be processed to obtain a standardized input image. S2. Input the standardized input image into the shared feature coding unit, perform multi-scale feature extraction, and output a shared coded feature map containing global facial structural features and local iris texture features; S3. Input the shared coded feature map into the face image enhancement branch and the iris recognition branch respectively, and perform parallel collaborative decoding. The face image enhancement branch completes adaptive enhancement of the face image based on the shared coded feature map and outputs the enhanced face image. The iris recognition branch completes iris region localization and iris feature extraction simultaneously based on the shared coded feature map and outputs the standard iris feature vector. S4. Calculate the image quality evaluation factor of the original image to be processed. Based on the image quality evaluation factor, adaptively adjust the fusion weight of the global facial features and the standard iris feature vector corresponding to the enhanced face image to generate a fused identity feature vector. Match the fused identity feature vector with the pre-stored registration feature library and output the identity recognition result.

2. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, The size normalization process in S1 specifically involves scaling the original image to be processed to a resolution of 640×480 pixels. The grayscale conversion uses a weighted average method, and the grayscale conversion formula is as follows: in, This represents the grayscale value of the pixel after grayscale conversion. , , These are the red, green, and blue channel component values ​​of the corresponding pixels in the original color image.

3. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, The shared feature encoding unit in step S2 adopts a lightweight MobileNetV2backbone architecture, which includes 4 downsampling modules and 8 inverse residual bottleneck modules. The first 2 downsampling modules extract global facial structural features, while the last 2 downsampling modules and 8 inverse residual bottleneck modules extract local iris texture features. The output shared encoded feature map has 128 channels and a feature map size of 40×30 pixels.

4. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, The face image enhancement branch in S3 includes a multi-scale feature fusion module and an upsampling decoding module. The optimization objective of the face image enhancement branch is a double-constraint loss function, with the following formula: in, For the total loss function, For pixel-level reconstruction loss, Preserve loss for iris texture , To balance the weighting coefficients, and , The value range is 0.4 to 0.

6. The value range is 0.4 to 0.

6.

5. The integrated iris recognition and face image enhancement processing method according to claim 4, characterized in that, The loss of iris texture preservation The calculation formula is: in, This is the function for calculating the mean square error. To enhance the feature vector of the iris region in the post-face image, This is the baseline feature vector corresponding to a standard clear iris image.

6. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, The iris recognition branch in S3 includes, in sequence, an iris region localization module, an iris normalization module, and an iris feature extraction module. The iris region localization module uses the Hough circle transform algorithm based on a shared coded feature map to accurately locate the inner and outer boundaries of the iris, with the localization deviation threshold controlled within 2 pixels. The iris normalization module expands the located annular iris region into a rectangular normalized iris image with a size of 512×64 pixels. The iris feature extraction module extracts a standard iris feature vector with a dimension of 256 based on the normalized iris image.

7. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, Image quality evaluation factors in S4 The calculation formula is: in, To score the uniformity of illumination, Rate the image sharpness. Scoring facial pose deviation. Score the degree of face occlusion. , , , For the corresponding weight coefficients, and , The range of values ​​is , A higher value indicates better image quality.

8. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, S4 is based on image quality evaluation factors The specific method for adaptively adjusting the fusion weights is as follows: Fusion weights of global facial features Fusion weights of iris feature vectors Fusion of identity feature vectors The calculation formula is: in, To enhance the 256-dimensional global feature vector of the face extracted from the post-face image, It is a 256-dimensional standard iris feature vector.

9. The integrated iris recognition and face image enhancement processing method according to claim 1, characterized in that, In S4, feature matching specifically involves calculating the cosine similarity between the fused identity feature vector and each registered feature vector in the registration feature library. When the maximum cosine similarity is greater than the preset similarity threshold of 0.85, the identity recognition result of successful matching is output; otherwise, the recognition result of failed matching is output.

10. An integrated iris recognition and face image enhancement processing system, using the integrated iris recognition and face image enhancement processing method as described in claims 1-9, characterized in that, include: The image preprocessing unit is used to acquire the original image to be processed, and to perform size normalization and grayscale preprocessing, and output a standardized input image; A shared feature coding unit is communicatively connected to the image preprocessing unit and is used to perform multi-scale feature extraction on the standardized input image and output a shared coded feature map. The dual-branch collaborative decoding unit, which is communicatively connected to the shared feature encoding unit, includes a face image enhancement branch module and an iris recognition branch module, which are used to perform adaptive face image enhancement and iris feature extraction in parallel, and output the enhanced face image and standard iris feature vector respectively. The quality-perceived fusion and recognition matching unit is communicatively connected to the dual-branch collaborative decoding unit and the image preprocessing unit, respectively. It is used to calculate the image quality evaluation factor, adaptively adjust the feature fusion weights and generate a fused identity feature vector, and complete the output of feature matching and identity recognition results.

Images