Face recognition method and system based on lens-free imaging
By constructing the lens-free face recognition network FAC-Net, the problem of accurately capturing face features from low-quality images obtained from lens-free imaging technology is solved, efficient face recognition in various application scenarios is achieved, and privacy protection capabilities are enhanced.
Patent Information
- Application Number
- CN202510132424.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-06
- Publication Date
- 2025-05-16
AI Technical Summary
The prior art cannot accurately capture face features from low-quality images obtained by lens-free imaging technology in a variety of application scenarios, and traditional face recognition methods are prone to causing privacy leakage problems.
Using a face recognition method based on lens-free imaging, the FAC-Net, a lens-free face recognition network, includes a multi-layer feature integration module and feature enhancement and denoising module, to extract and enhance face features to achieve accurate recognition of low-quality images.
In a variety of application scenarios, precise capture of face features is achieved, reducing the risk of privacy leakage, and improving the accuracy and robustness of face recognition.
Smart Images
Figure CN120014686A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition, and in particular to a face recognition method and system based on lens-free imaging. Background Art
[0002] Traditional lens-based imaging technology directly captures individual facial features, which has led to growing concerns about privacy leaks. As an innovative image acquisition method, lens-free imaging technology uses scattering elements such as scattering plates and frosted glass to replace traditional lenses, effectively avoiding the risk of directly capturing individual facial details and significantly reducing the possibility of personal information leaks.
[0003] In the prior art, traditional face recognition methods directly capture and process user video images, which involves the user's facial and body features and is prone to privacy leakage. In particular, when cameras are installed in public places or private spaces for sitting posture monitoring, the user's privacy is difficult to guarantee. In addition, video imaging is sensitive to lighting conditions. Insufficient light or the presence of obstructions will seriously affect the image quality, thereby reducing the accuracy of face recognition.
[0004] In summary, existing technologies are unable to accurately capture facial features from low-quality images obtained by lensless imaging technology in a variety of application scenarios. Summary of the invention
[0005] The embodiments of the present invention provide a face recognition method and system based on lens-free imaging, which can solve the problem in the prior art that it is impossible to accurately capture facial features from low-quality images obtained from lens-free imaging technology in various application scenarios.
[0006] The embodiment of the present invention provides a face recognition method based on lens-free imaging, comprising the following steps: acquiring a face image recognized by the lens-free imaging technology; constructing a lens-free face recognition network FAC-Net, wherein the FAC-Net comprises: a multi-layer feature integration module and a feature enhancement and denoising module; wherein the multi-layer feature integration module comprises: the second, third and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module comprises: convolution kernels of multiple scales and residual connections; training the lens-free face recognition network FAC-Net to recognize faces through face images, and obtaining a training The trained FAC-Net is input into the trained FAC-Net, and in the multi-layer feature integration module, the facial features are extracted at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50 respectively; the downsampling results of the second and third residual blocks are superimposed with the upsampling result of the fourth residual block to obtain facial features with increased detail recognition; the facial features with increased detail recognition are input into the feature enhancement and denoising module, and local features of multiple scales are extracted through convolution kernels of multiple scales, and the local feature maps of multiple scales are added to the facial features with increased detail recognition element by element through residual connection to obtain facial features with increased image quality; and face recognition is completed.
[0007] Furthermore, the lens-free face recognition network FAC-Net also includes: a cascade grouping attention module.
[0008] Furthermore, the facial features that increase the image quality need to be input into a cascaded grouping attention module for processing; the processing steps include: inputting the facial features that increase the image quality into a cascaded grouping attention module used to extract feature information of different regions, focusing on features of different facial regions through the attention mechanisms of different groups, and obtaining facial recognition results that capture feature details of different facial regions.
[0009] Furthermore, the step of obtaining the facial features for adding detail recognition comprises: In ResNet50, the second residual block contains 4 residual units, the output channel of the first 1×1 convolution layer is 128, and the output channel of the fourth 1×1 convolution layer is 512; the third residual block contains 6 residual units, the output channel of the first 1×1 convolution layer is 256, and the output channel of the sixth 1×1 convolution layer is 1024; the fourth residual block contains 3 residual units, the output channel of the first 1×1 convolution layer is 512, and the output channel of the sixth 1×1 convolution layer is 2048; The 2 residual blocks, the 3rd residual block and the 4th residual block extract facial features at different levels respectively; the facial features extracted by the 2nd residual block and the 3rd residual block are downsampled respectively to obtain a first feature map and a second feature map, and the first feature map and the second feature map have the same size; the facial features extracted by the 4th residual block are upsampled to obtain a third feature map, and the third feature map has the same size as the first feature map and the second feature map; the first feature map, the second feature map and the third feature map are added element by element to obtain facial features for increased detail recognition.
[0010] Furthermore, the facial features for increasing image quality are obtained by specifically inputting the facial features for increasing detail recognition into a feature enhancement and denoising module, capturing multi-level facial features through 3×3, 5×5, 7×7 and 9×9 convolution kernels respectively; and adding the multi-level facial features to the facial features for increasing detail recognition element by element to obtain the facial features for increasing image quality.
[0011] Furthermore, the method of obtaining a face recognition result that captures the feature details of different facial regions comprises the following specific steps: dividing the facial features that increase the image quality into multiple groups according to the spatial regions; obtaining the attention weight of each group, and weighting the features according to the attention weight; performing attention processing on each group and cascading all the groups, wherein the cascade uses the output of the attention processing of each group as the input of the attention processing of the next group; performing a linear transformation on the cascade result using a linear layer to obtain a feature vector, and inputting the feature vector into a classifier to obtain a face recognition result that captures the feature details of different facial regions.
[0012] Furthermore, the method of obtaining a facial image recognized by the lens-free imaging technology comprises the following specific steps: removing the lens of the Raspberry Pi PiCamera to obtain a sensor; surrounding the sensor with the prepared separation material, and sealing the sides of the sensor with black tape to ensure that light does not affect the sensor, thereby completing the production of the lens-free camera; installing the lens-free camera on the Raspberry Pi to form a recognition device, and installing the neural network RLCNet on the Raspberry Pi; using the facial image to train the neural network RLCNet to recognize the face, thereby obtaining a facial recognition model; and inputting the image obtained by the lens-free camera into the facial recognition model, thereby obtaining a facial image.
[0013] An embodiment of the present invention provides a face recognition system based on lens-free imaging, comprising: An image acquisition module is used to acquire a face image recognized by the lens-free imaging technology; a FAC-Net construction module is used to construct a lens-free face recognition network FAC-Net, wherein the FAC-Net includes: a multi-layer feature integration module and a feature enhancement and denoising module; wherein the multi-layer feature integration module includes: the second, third and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module includes: convolution kernels of multiple scales and residual connections; the lens-free face recognition network FAC-Net is trained to recognize faces through face images to obtain the trained FAC-Net et; a face recognition module, which is used to input the face image to be tested into the trained FAC-Net, and in the multi-layer feature integration module, extract face features at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50; superimpose the downsampling results of the second and third residual blocks with the upsampling result of the fourth residual block to obtain face features with increased detail recognition; input the face features with increased detail recognition into the feature enhancement and denoising module, extract local features of multiple scales through convolution kernels of multiple scales, and add the local feature maps of multiple scales and the face features with increased detail recognition element by element through residual connections to obtain face features with increased image quality; and complete face recognition.
[0014] The embodiment of the present invention provides a face recognition method and system based on lens-free imaging, which has the following beneficial effects compared with the prior art: A lens-free face recognition network FAC-Net is constructed, including: a multi-layer feature integration module and a feature enhancement and denoising module; the face image to be tested is input into the trained FAC-Net, and in the multi-layer feature integration module, the face features are extracted at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50 respectively; the downsampling results of the second and third residual blocks are superimposed with the upsampling result of the fourth residual block to obtain face features for increased detail recognition; the face features for increased detail recognition are input into the feature enhancement and denoising module, local features of multiple scales are extracted through convolution kernels of multiple scales, and local feature maps of multiple scales are added to the face features for increased detail recognition through residual connections to obtain face features for increased image quality; and face recognition is completed.
[0015] Finally, faced with low-quality facial images obtained by lens-free imaging technology in various application scenarios, facial features are extracted at different levels through a multi-layer feature integration module, and the downsampling results of the second and third residual blocks are superimposed with the upsampling results of the fourth residual block to obtain facial features with increased detail recognition. On the basis of the multi-layer feature integration module focusing on facial features and ignoring the remaining features in various application scenarios, the feature enhancement and denoising module is used to further perform local feature extraction and denoising on the facial features, thereby achieving accurate capture of facial features in various application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 The overall architecture diagram of FAC-Net provided by the embodiment of the present invention; Figure 2 A schematic diagram of a multi-layer feature integration module provided by an embodiment of the present invention; Figure 3 A schematic diagram of a feature enhancement and denoising module provided in an embodiment of the present invention; Figure 4 A schematic diagram of a cascaded grouping attention module provided in an embodiment of the present invention; Figure 5 A flowchart of a face recognition method based on lens-free imaging provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.
[0018] The embodiment of the present invention provides a face recognition method based on lens-free imaging, comprising the following steps: Step 1: Obtain a face image recognized by lens-free imaging technology.
[0019] Step 2: Construct a lens-free face recognition network FAC-Net, wherein the FAC-Net includes: a multi-layer feature integration module, a feature enhancement and denoising module, and a cascade group attention module; wherein the multi-layer feature integration module includes: the second, third, and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module includes: convolution kernels of multiple scales and residual connections; the lens-free face recognition network FAC-Net is trained to recognize faces through face images to obtain a trained FAC-Net.
[0020] Step 3: Input the face image to be tested into the trained FAC-Net, and in the multi-layer feature integration module, extract face features at different levels through the 2nd, 3rd and 4th residual blocks of the residual layer in ResNet50; superimpose the downsampling results of the 2nd and 3rd residual blocks with the upsampling result of the 4th residual block to obtain face features with increased detail recognition; input the face features with increased detail recognition into the feature enhancement and denoising module, extract local features of multiple scales through convolution kernels of multiple scales, and add the local feature maps of multiple scales and the face features with increased detail recognition element by element through residual connections to obtain face features with increased image quality; input the face features with increased image quality into the cascade grouping attention module used to extract feature information of different regions, focus on the features of different facial regions through the attention mechanisms of different groups, obtain face recognition results that capture the feature details of different facial regions, and complete face recognition.
[0021] 1. Production of lens-less camera.
[0022] Remove the lens from the Raspberry Pi Camera and use a razor blade to cut the glue between the bracket and the sensor. When cutting, be careful not to damage any electrical components on the edge of the sensor to determine the location of the delicate components, and then start cutting from the side with fewer components.
[0023] Place the separator material (cardboard, about 3mm thick) around the sensor, using black tape to make sure the separator doesn't let light in from the sides, then place double-sided tape on top of it to keep it taut and supported at the correct focal distance.
[0024] A lensless camera is mounted on the Raspberry Pi to form a recognition device, and a neural network RLCNet model is installed on the Raspberry Pi to recognize faces.
[0025] 8. The overall structure of the lens-free face recognition network FAC-Net network model.
[0026] In order to deal with the problems of missing image feature information, noise and distortion in lens-free imaging, the present invention innovatively proposes the FAC-Net model, a lens-free face recognition network that integrates stacked feature extraction and regional attention enhancement. The model is based on ResNet-50. Through the multi-level feature integration module (Multi-Level Feature Integration Module, MFIM) fine convolution and stacked fusion of different levels of features, the model enhances the capture of subtle facial features in blurry and low-quality images. Subsequently, the Feature Enhancement and Denoising Module (Feature Enhancement and Denoising Module, FED) can not only effectively suppress noise and redundant information while retaining key facial features, but also improve image recognizability through advanced feature enhancement techniques such as sharpening and deblurring. Finally, the Cascaded Group Attention (CGA) is introduced to intelligently focus on key facial areas, improve the model's adaptability to complex feature changes, and ensure that face recognition in lens-free environments is both efficient and accurate. Figure 1 shown.
[0027] 3. Composition of the FAC-Net network model.
[0028] 1) Multi-layer feature integration module (MFIM): In order to more effectively address the challenges of limited image quality and reduced information in lensless imaging, we carefully designed a multi-layer feature integration module (MFIM). This module uses the deep learning capabilities of convolutional neural networks (CNNs) to finely extract feature maps at different levels. Subsequently, through the feature integration layer, feature maps from different depth levels are cleverly upsampled, downsampled, and superimposed to achieve comprehensive integration and enhancement of features. This process ensures that the model can capture rich feature information in images from multiple scales and depths, thereby significantly improving the accuracy and robustness of sitting posture recognition.
[0029] Specifically, the MFIM module focuses on the three key levels of STAGE2, STAGE3, and STAGE4 in the ResNet50 network, and deeply integrates the features extracted from each of them. This cross-level feature integration strategy aims to make full use of the diverse features that exist widely from shallow to deep layers, so that the model can integrate multi-level contextual information during the feature extraction process and achieve a deeper understanding of the image content. During the integration process, we carefully designed the algorithm to retain key feature information while optimizing the feature dimension and improving computational efficiency.
[0030] The introduction of the MFIM module not only significantly enhances the model's ability to extract image features, but also effectively alleviates the problem of insufficient learning ability of shallow networks commonly seen in ResNet by promoting shallow networks to receive more gradient information during training. Figure 2 As shown in Figure 2, this module ensures that all layers of the model can fully participate in the learning process, thereby optimizing the overall performance. Through the ingenious design of the MFIM module, we provide strong technical support for the sitting posture recognition task under lens-free imaging conditions.
[0031] 2) Feature enhancement and denoising module (FED): In lensless imaging technology, since it does not rely on traditional optical lenses to focus and guide light, but directly obtains image information through scattering elements, this process is easily affected by multiple factors such as light scattering, diffraction and environmental interference, resulting in image noise. To this end, the present invention introduces a feature enhancement and denoising module (FED), which uses a multi-scale analysis strategy to deeply observe and process images at multiple scales to effectively distinguish and enhance useful signals in the image (such as edges, textures and other detail information) while suppressing noise.
[0032] The core of the FED module lies in its multi-scale processing capability, which allows the system to capture structures and details of different sizes in images at different scales. In this framework, high-frequency components, as important carriers of image details, are rich in edge and texture information, but are also easily mixed with noise. The FED module uses a carefully designed algorithm to decompose and reconstruct images at different scales, thereby accurately distinguishing between signals and noise, and achieving feature enhancement and noise suppression.
[0033] To achieve this goal, the FED module uses convolution kernels of various scales, specifically 3x3, 5x5, 7x7 and 9x9, to fully capture multi-level features from fine edges to overall textures. In the feature extraction stage, these convolution kernels act on the input feature map respectively, extracting local features at their respective scales through convolution operations, forming a series of feature maps containing information at different levels of detail.
[0034] Then, in the feature enhancement and denoising stage, the FED module adopts a clever fusion strategy. It adds the convolution results at each scale to the original feature map element by element. This operation not only retains the key information in the original image, but also enhances the model's perception of image details by introducing multi-scale features. At the same time, due to the complementarity of convolution kernels of different scales, this process also indirectly suppresses noise and improves the overall quality of the image.
[0035] In summary, the FED module effectively solves the problems of image noise and blurred details in lensless imaging through its multi-scale analysis, feature extraction and fusion strategy, and significantly improves the quality of image features and the recognition performance of the model. Figure 3 shown.
[0036] 3) Cascaded Group Attention (CGA): Cascaded Group Attention (CGA) is an attention mechanism used to improve the computational efficiency and performance of the model. The core idea is to provide different input splits for each attention head and cascade the output features across heads, thereby enhancing the diversity of features input to the attention heads. Specifically, CGA divides the input features into different parts, each of which is input to an attention head. Each head calculates its self-attention map, then concatenates the outputs of all heads and projects them back to the dimension of the input through a linear layer. In this way, CGA improves the computational efficiency of the model without adding additional parameters. In addition, through concatenation, the output of each head is added to the input of the next head, thereby gradually refining the feature representation. CGA not only reduces computational redundancy in multi-head attention, but also improves model capacity by increasing network depth. Since the QK channel dimension of each head is small, only a small delay overhead is added. CGA modules such as Figure 4 shown.
[0037] 4. The privacy-preserving face recognition technology based on lens-free imaging proposed in the present invention is mainly dedicated to solving the following key technical problems.
[0038] 1) Enhanced privacy protection: Traditional lens-based face recognition technology directly captures and processes individual facial details, which can easily lead to the leakage of personal privacy. This technology adopts a lens-free imaging method, replacing traditional lenses with scattering elements, avoiding the direct generation of clear facial images, thereby greatly reducing the risk of exposure of personal identity information at the source. This technology realizes the face recognition function while ensuring the privacy of users.
[0039] 2) Overcoming image quality challenges: Due to the scattering effect, lensless imaging technology often leads to a significant decrease in image quality, which is manifested as image blur and increased noise. This poses a severe challenge to accurately extracting facial features from images. To solve this problem, the present invention has developed advanced image processing algorithms and deep learning models to effectively deal with blur and noise, and achieve accurate capture of facial features in low-quality images.
[0040] 3) Optimizing feature extraction and recognition accuracy: In a lensless imaging environment, due to image quality limitations, traditional facial feature extraction and recognition algorithms are difficult to apply directly. To this end, this paper introduces innovative modules such as multi-layer feature fusion, feature enhancement, and regional attention mechanism, aiming to effectively extract and enhance key facial features from blurred images, significantly improving the accuracy and robustness of face recognition.
[0041] 4) Improve model robustness and generalization ability: In practical applications, lensless imaging systems may encounter a variety of complex environments and conditions, such as changes in lighting conditions and increased background complexity. To meet these challenges, this paper significantly improves the robustness and generalization ability of the model through a carefully designed model structure and optimized training strategy, ensuring that it can maintain stable recognition performance in different application scenarios.
[0042] To sum up, the privacy-preserving face recognition technology based on lens-free imaging proposed in the present invention not only effectively solves the problem of how to accurately extract facial features from low-quality lens-free imaging images while protecting personal privacy, but also improves the recognition accuracy, robustness and generalization ability of the model through technological innovation, providing a new solution for the combination of privacy protection and efficient face recognition technology.
[0043] 5. The effects produced by the present invention.
[0044] 1) Significant privacy protection effect: Avoid direct facial capture: This technology avoids the problem of traditional cameras directly capturing user facial features through lens-free imaging, significantly reduces the risk of facial privacy leakage, and enhances user privacy protection.
[0045] Facial information blurring: The speckle pattern generated by lensless imaging technology highly blurs the direct visual features of the original facial image while retaining the necessary spatial information, making it difficult to easily identify and restore sensitive facial information.
[0046] 2) Improved face recognition accuracy and robustness: High-precision recognition: With the help of advanced image processing algorithms and deep learning models, this technology can accurately extract facial features from blurry, noisy lens-free imaging images to achieve high-accuracy face recognition.
[0047] Strong robustness: By continuously optimizing the algorithm and model structure, this technology can maintain stable recognition performance in complex scenarios such as different lighting conditions, facial occlusion, and expression changes, demonstrating strong robustness.
[0048] 3) User experience optimization: Contactless recognition: Users can achieve face recognition without any contact or wearing any equipment, which lowers the usage threshold and improves the convenience and comfort of use.
[0049] Wide applicability: This technology can be widely used in various scenarios such as home, office, public places, etc., providing users with convenient and efficient face recognition services to meet diverse needs.
[0050] 4) Promote technology application and promotion: Expanding application areas: Given its strong privacy protection capabilities, high-precision recognition and good user experience, this technology has broad application prospects and promotion value, and can be applied to multiple fields such as security access control, payment verification, and personalized services.
[0051] Promote technological innovation: The successful application of this technology will further promote in-depth research and innovation in the field of lens-free imaging technology and deep learning algorithms in the field of face recognition, and promote the continuous advancement of related technologies.
[0052] 5) Significant social benefits: Enhance public safety: Through efficient and accurate face recognition technology, it helps to improve the level of security monitoring in public places and enhance the overall public safety of society.
[0053] Improve the convenience of life: In daily scenarios such as payment and access control, this technology can simplify the operation process, improve the user experience, and bring more convenience to people's lives. At the same time, it also helps to reduce the inconvenience and security risks caused by identity verification problems.
[0054] An embodiment of the present invention provides a face recognition system based on lens-free imaging, comprising: The image acquisition module is used to acquire the face image recognized by the lens-free imaging technology.
[0055] The FAC-Net building module is used to build a lens-free face recognition network FAC-Net, and the FAC-Net includes: a multi-layer feature integration module and a feature enhancement and denoising module; wherein the multi-layer feature integration module includes: the second, third and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module includes: convolution kernels and residual connections of multiple scales; the lens-free face recognition network FAC-Net is trained to recognize faces through face images to obtain a trained FAC-Net.
[0056] The face recognition module is used to input the face image to be tested into the trained FAC-Net, and in the multi-layer feature integration module, extract the face features at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50; superimpose the downsampling results of the second and third residual blocks with the upsampling result of the fourth residual block to obtain the face features with increased detail recognition; input the face features with increased detail recognition into the feature enhancement and denoising module, extract local features of multiple scales through convolution kernels of multiple scales, and add the local feature maps of multiple scales and the face features with increased detail recognition element by element through residual connection to obtain the face features with increased image quality; and complete face recognition.
[0057] A specific embodiment is as follows: An application of intelligent door opening system based on FAC-Net network model: (9) When the user approaches the access control system, the system activates the lensless imaging module to collect the speckle pattern on the user's face. The image processing module then pre-processes the collected speckle pattern to improve the image quality.
[0058] (10) Then, the FAC-Net network model recognition module extracts and recognizes features of the preprocessed image and outputs the recognition results.
[0059] (11) Finally, the access control module intelligently controls the switch of the access control system based on the recognition results, enabling users to conveniently open the door.
[0060] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A face recognition method based on lens-free imaging, characterized in that: The following steps are involved: Acquire a face image recognized by lens-free imaging technology; A lens-free face recognition network FAC-Net is constructed, wherein the FAC-Net comprises: a multi-layer feature integration module and a feature enhancement and denoising module; wherein the multi-layer feature integration module comprises: the second, third and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module comprises: convolution kernels of multiple scales and residual connections; The lens-free face recognition network FAC-Net is trained through face images to recognize faces, and the trained FAC-Net is obtained; The face image to be tested is input into the trained FAC-Net, and in the multi-layer feature integration module, the face features are extracted at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50 respectively; the downsampling results of the second and third residual blocks are superimposed with the upsampling result of the fourth residual block to obtain the face features with increased detail recognition; the face features with increased detail recognition are input into the feature enhancement and denoising module, and local features of multiple scales are extracted through convolution kernels of multiple scales, and the local feature maps of multiple scales are added to the face features with increased detail recognition element by element through residual connection to obtain the face features with increased image quality; and face recognition is completed.
2. A face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The lens-free face recognition network FAC-Net also includes: a cascade grouping attention module.
3. The face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The facial features that increase the image quality need to be input into the cascade grouping attention module for processing; The processing steps include: inputting facial features that increase image quality into a cascade grouping attention module for extracting feature information of different regions, focusing on features of different facial regions through attention mechanisms of different groups, and obtaining facial recognition results that capture feature details of different facial regions.
4. The face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The specific steps of obtaining the facial features for adding detail recognition include: In ResNet50, the second residual block contains 4 residual units, the output channel of the first 1×1 convolution layer is 128, and the output channel of the fourth 1×1 convolution layer is 512; The third residual block contains 6 residual units, the output channel of the first 1×1 convolution layer is 256, and the output channel of the sixth 1×1 convolution layer is 1024; The fourth residual block contains 3 residual units, the output channel of the first 1×1 convolution layer is 512, and the output channel of the sixth 1×1 convolution layer is 2048; The second residual block, the third residual block, and the fourth residual block in ResNet50 are used to extract facial features at different levels. Down-sampling the facial features extracted by the second residual block and the third residual block respectively to obtain a first feature map and a second feature map, and the first feature map and the second feature map have the same size; Upsampling the facial features extracted by the fourth residual block to obtain a third feature map, and the third feature map has the same size as the first feature map and the second feature map; The first feature map, the second feature map, and the third feature map are added element by element to obtain facial features for increased detail recognition.
5. The face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The steps of obtaining facial features that increase image quality include: The facial features with added detail recognition are input into the feature enhancement and denoising module, and multi-level facial features are captured through 3×3, 5×5, 7×7 and 9×9 convolution kernels respectively; The multi-level facial features are respectively added element by element with the facial features for increasing detail recognition to obtain the facial features for increasing image quality.
6. The face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The step of obtaining a face recognition result that captures the feature details of different facial regions comprises: Dividing facial features that increase image quality into multiple groups according to spatial regions; Get the attention weight of each group and weight the features according to the attention weight; Performing attention processing on each group and cascading all the groups, wherein the cascade is to use the output of the attention processing of each group as the input of the attention processing of the next group; The cascade result is linearly transformed using a linear layer to obtain a feature vector, which is then input into a classifier to obtain a face recognition result that captures the feature details of different facial regions.
7. The face recognition method based on lens-free imaging as claimed in claim 1, characterized in that: The step of obtaining a face image recognized by the lens-free imaging technology specifically comprises: Remove the lens of the Raspberry Pi Camera to get the sensor; Place the prepared separation material around the sensor and seal the sides of the sensor with black tape to ensure that light does not affect the sensor. This completes the production of the lensless camera. A lensless camera is mounted on the Raspberry Pi to form a recognition device, and the neural network RLCNet is installed on the Raspberry Pi; Use face images to train the neural network RLCNet to recognize faces and obtain a face recognition model; The image obtained by the lensless camera is input into the face recognition model to obtain a face image.
8. A face recognition system based on lens-free imaging, characterized in that: include: An image acquisition module, used to acquire a face image recognized by the lens-free imaging technology; A FAC-Net construction module is used to construct a lens-free face recognition network FAC-Net, wherein the FAC-Net includes: a multi-layer feature integration module and a feature enhancement and denoising module; wherein the multi-layer feature integration module includes: the second, third and fourth residual blocks of the residual layer in the neural network ResNet50; downsampling is used for the second and third residual blocks, and upsampling is used for the fourth residual block; the feature enhancement and denoising module includes: convolution kernels of multiple scales and residual connections; the lens-free face recognition network FAC-Net is trained to recognize faces through face images to obtain a trained FAC-Net; The face recognition module is used to input the face image to be tested into the trained FAC-Net, and in the multi-layer feature integration module, extract the face features at different levels through the second, third and fourth residual blocks of the residual layer in ResNet50; superimpose the downsampling results of the second and third residual blocks with the upsampling result of the fourth residual block to obtain the face features with increased detail recognition; input the face features with increased detail recognition into the feature enhancement and denoising module, extract local features of multiple scales through convolution kernels of multiple scales, and add the local feature maps of multiple scales and the face features with increased detail recognition element by element through residual connections to obtain the face features with increased image quality; and complete face recognition.