Face recognition method and system based on dynamic weighted wavelet attention

Abstract

The invention discloses a face recognition method and system based on dynamic weighted wavelet attention, and the method comprises the steps: inputting a face image into a backbone network; extracting multi-scale feature vectors and adjusting feature intensity by using a spatial depth conversion convolution layer, a feature extraction enhancement module and a spatial pyramid pooling fast layer in the backbone network; a detail feature enhancement module is introduced into a feature extraction enhancement module, high-frequency information of an image is extracted through wavelet transform, and then detail features are enhanced through channel attention and space attention, so that the expression ability of textures, edges and other details is enhanced while space structure information of middle features is reserved. Multi-scale feature vectors extracted by a backbone network are input into a neck network, feature vectors after feature fusion are output and enter a detection layer, face recognition is performed by using detection heads of different sizes, and the small target face detection precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of face recognition, and in particular relates to a face recognition method and system based on dynamic weighted wavelet attention. Background Art

[0002] Object detection has long been a core problem in computer vision research. Its mission is to distinguish objects from other regions of interest in images or videos, determine their presence, and confirm their category and location. With the widespread application of feature extraction techniques based on deep neural networks in computer vision tasks, object detection technology has achieved significant breakthroughs. Currently, object detection algorithms are generally categorized into two-stage and one-stage algorithms, depending on the rules for generating candidate regions.

[0003] The former primarily uses deep learning methods to extract deep features from candidate regions. Representative algorithms include R-CNN, SPP-net, Fast R-CNN, Faster R-CNN, and R-FCN. Although two-stage algorithms maintain a significant advantage in accuracy, their detection speed is often unsatisfactory in real-time scene detection.

[0004] In contrast, the latter method does not require the selection of candidate target regions, but directly regresses the target's distribution probability and position coordinates, significantly improving detection speed. Due to their advantages of high detection accuracy and reduced time consumption, single-stage algorithms have gradually become a research hotspot in the field of target detection. Representative algorithms include RetinaNet, the YOLO series, and the SSD series. However, existing target detection methods have low accuracy when detecting small targets. In the field of face recognition, this is particularly problematic due to low accuracy, false detections, and missed detections. Therefore, a face recognition model is needed that can quickly and accurately identify small target faces in complex environments. Summary of the Invention

[0005] The purpose of the present invention is to provide a face recognition method and system based on dynamic weighted wavelet attention.

[0006] In a first aspect, the present invention provides a face recognition method based on dynamic weighted wavelet attention, comprising: Construct a dataset containing facial images and preprocess the facial images in the dataset; construct a face recognition model; the face recognition model includes a backbone network, a neck network, and a detection head; the backbone network is used to extract multi-level features from the input image and input them into the neck network for fusion; the detection head is used to convert the fused features into recognition results; The backbone network introduces a detail feature enhancement module to extract detail features; in the detail feature enhancement module, the input feature map Decompose into low-frequency components and high-frequency components, and divide the high-frequency components into horizontal high-frequency features, vertical high-frequency features and diagonal high-frequency features; splice the horizontal high-frequency features, vertical high-frequency features and diagonal high-frequency features to obtain the overall high-frequency features; enhance the overall high-frequency features through the convolution block, and reconstruct the low-frequency components and the enhanced high-frequency features to obtain the enhanced feature map; compare the enhanced feature map with the feature map of the input detail feature enhancement module Perform element-by-element addition and fusion to obtain a fused feature map; process the fused feature map through the channel attention mechanism and the spatial attention mechanism respectively to obtain the channel weight and spatial weight; multiply the fused feature map with the channel weight and the spatial weight successively to obtain the feature map output by the feature extraction enhancement module; perform face recognition on the tested image by introducing the face recognition model of the detail feature enhancement module.

[0007] Preferably, the channel attention mechanism performs adaptive average pooling on the fused feature map to extract a global description of the channel direction; the global description is processed by two convolutional networks to obtain channel weights.

[0008] Preferably, the spatial attention mechanism performs a convolution operation on the fused feature map to obtain a spatial weight.

[0009] Preferably, the backbone network includes multiple feature extraction layers and a spatial pyramid pooling fast layer connected in sequence. The input image is input into the feature extraction layer after convolution processing for feature extraction, and the feature map processed by the multiple feature extraction layers is input into the spatial pyramid pooling fast layer; the feature maps output by the spatial pyramid pooling fast layer and each feature extraction layer are used as the feature maps output by the backbone network.

[0010] Preferably, the feature extraction layer includes a spatial depth conversion convolution layer and a feature extraction enhancement module connected in sequence; the spatial depth conversion convolution layer is used to reorganize the spatial data of the input feature map into channel data, and perform channel compression on the reorganized result through a non-step convolution operation; The feature extraction enhancement module processes the feature map output by the spatial depth conversion convolution layer through a convolution block to obtain an intermediate feature map; splits the intermediate feature map into two feature maps, one of which is processed in turn by multiple bottleneck modules, and the feature map output by each bottleneck module is spliced ​​with the other feature map in the channel dimension; the spliced ​​result is finally transformed through a convolution block to obtain a feature map output by the feature extraction enhancement module.

[0011] Preferably, the bottleneck module includes two convolution blocks connected in sequence and a detail feature enhancement module; the result of splicing the feature map output by the detail feature enhancement module and the feature map input to the bottleneck module is used as the feature map output by the bottleneck module; the two convolution blocks have the same structure, both including a convolution layer, a batch normalization layer and an activation function connected in sequence.

[0012] Preferably, the neck network includes a feature pyramid network and a path aggregation network; the feature pyramid network includes a plurality of semantic feature fusion layers connected in sequence; the path aggregation network includes a plurality of detail feature fusion layers connected in sequence.

[0013] Preferably, in the feature pyramid network, each semantic feature fusion layer processes the input feature map by upsampling, and adds the processed feature map to the feature map output by the corresponding feature extraction layer element by element, and then processes it through the C2f module to obtain the feature map output by the semantic feature fusion layer; except for the first semantic feature fusion layer, the feature map input to each semantic feature fusion layer is the feature map output by the previous semantic feature fusion layer; the feature map input to the first semantic feature fusion layer is the feature map output by the spatial pyramid pooling fast layer; In the path aggregation network, except for the last detail feature fusion layer, each detail feature fusion layer processes the input feature map through convolution operation, and adds the processed feature map to the feature map output by the corresponding semantic feature extraction layer element by element. After processing by the C2f module, the feature map output by the detail feature fusion layer is obtained; after convolution processing, the last detail feature fusion layer fuses the feature map output by the spatial pyramid pooling fast layer; except for the first detail feature fusion layer, the feature map input to each detail feature fusion layer is the feature map output by the previous detail feature fusion layer; the feature map input to the first detail feature fusion layer is the feature map output by the last semantic feature extraction layer.

[0014] Preferably, the preprocessing process is: using a mosaic data enhancement method to process the face images in the data set, and performing random scaling, cropping and arrangement.

[0015] In the second aspect, the present invention provides a face recognition system based on dynamic weighted wavelet attention, which is used to execute the above-mentioned face recognition method; the face recognition system includes an image acquisition module, a preprocessing module and a face recognition module; the image acquisition module is used to acquire image data containing faces and input it into the preprocessing module for processing; the face recognition module is used to identify faces in images.

[0016] The present invention has the following beneficial effects: 1. The present invention introduces a detail feature enhancement module into the feature extraction layer, so that the extracted feature map can effectively enhance the expression of detail features such as texture and edge while retaining spatial structure information. At the same time, it improves the representation ability of key areas in the feature map, so that the feature extraction layer can show better performance in fine-grained feature extraction, solving the problem of insufficient learning of small targets due to the loss of fine-grained information in existing target recognition methods.

[0017] 2. The present invention achieves feature dimensionality reduction by adopting a spatial depth conversion convolution layer, and retains discriminative feature information through learnable convolution parameters; compared with traditional downsampling methods, the spatial depth conversion convolution layer avoids directly discarding pixel information and can better extract the feature details of small target faces. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of some facial images in the data set of the present invention.

[0019] Figure 2 Schematic diagram of the face recognition model structure in the present invention.

[0020] Figure 3 Schematic diagram of the spatial depth conversion convolution layer structure in the present invention.

[0021] Figure 4 This is a structural diagram of the feature extraction enhancement module in the present invention.

[0022] Figure 5 This is a schematic diagram of the bottleneck module structure in the present invention.

[0023] Figure 6 Schematic diagram of the detail feature enhancement module structure in the present invention.

[0024] Figure 7 Schematic diagram of the recognized face image output by the detection head in the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described below with reference to the accompanying drawings.

[0026] A face recognition method based on dynamic weighted wavelet attention uses a face recognition system consisting of an image acquisition module, a preprocessing module, and a face recognition module. The image acquisition module is used to collect image data containing faces and input it into the preprocessing module for processing; the face recognition module is used to identify faces in images.

[0027] The face recognition method comprises the following steps: Step 1: Build a dataset The dataset uses the PartA subset of SCUT-HEAD, which is a large-scale head detection dataset. PartA includes 2,000 images sampled from surveillance videos of a university classroom, such as Figure 1 As shown in the figure, 67,321 heads were annotated, covering scenes with complex lighting, occlusion, and small objects. Information such as the center coordinates, width, and height of the face object in each image was recorded. The images in the dataset were processed using the Mosaic data augmentation method, which randomly scaled, cropped, and permuted them, addressing the imbalance between small, medium, and large objects in the dataset.

[0028] Step 2: Build a face recognition model like Figure 2 As shown in the figure, the face recognition model includes a backbone network, a neck network, and a detection head. The backbone network is used to extract multi-scale feature vectors of the image. It includes four layers of feature extraction layers and a spatial pyramid pooling fast layer (SPPF) connected in sequence. The feature map input to the backbone network is input into the feature extraction layer after convolution processing for feature extraction. The feature map processed by the four layers of feature extraction layers is input into the spatial pyramid pooling fast layer, which improves the expressive power of the feature map. The neck network includes a feature pyramid network (FPN) and a path aggregation network (PAN); the feature pyramid network includes three semantic feature fusion layers connected in sequence; the path aggregation network includes three detail feature fusion layers connected in sequence; In the Feature Pyramid Network, each semantic feature fusion layer processes its input feature map through upsampling and element-by-element addition of the processed feature map to the feature map output by the corresponding (same-scale) feature extraction layer. After further processing by the C2f module, the output feature map of the semantic feature fusion layer is obtained. Except for the first semantic feature fusion layer, the input feature map of each semantic feature fusion layer is the feature map output by the previous semantic feature fusion layer; the input feature map of the first semantic feature fusion layer is the feature map output by the fast spatial pyramid pooling layer.

[0029] In the path aggregation network, except for the last detail feature fusion layer, each detail feature fusion layer processes the input feature map through a convolution operation. The processed feature map is then element-wise added to the feature map output by the corresponding (same-scale) semantic feature extraction layer. After processing through the C2f module, the output feature map of the detail feature fusion layer is obtained. After the convolution operation, the last detail feature fusion layer fuses the feature map output by the spatial pyramid pooling fast layer. Except for the first detail feature fusion layer, the input feature map of each detail feature fusion layer is the feature map output by the previous detail feature fusion layer. The input feature map of the first detail feature fusion layer is the feature map output by the last semantic feature extraction layer.

[0030] The feature maps output by each detail feature fusion layer and the final semantic feature fusion layer are fed into four detection heads of different sizes to detect faces in the image. For an input image of 640×640, a single grid in the feature map covers an area of ​​8×8 pixels. This makes it difficult for the network to extract features for objects smaller than 8 pixels in height or width, leading to missed detections and significantly negatively impacting small object face recognition tasks. A new 160×160 detection feature map is added to detect small objects larger than 4×4 and prevent information about small objects from being lost during deep network processing.

[0031] During object detection, the detection accuracy of small objects is often significantly lower than that of normal objects. Small object pixels make up a relatively small proportion of the entire image, leaving minimal background information for the model to learn from. Furthermore, small objects are often accompanied by other larger objects, which often dominate the learning process, resulting in the small objects being undetectable. In this situation, the reliability of convolutional neural networks is significantly reduced. In the early layers of a convolutional neural network architecture, image resolution is suitable for studying large objects, and stride volume effectively filters out redundant information, enabling the model to learn more effective features. However, when the image pixel count is low or the objects being detected are small, the amount of redundant data is less. In this case, strided convolution and pooling can lead to the loss of fine-grained information, resulting in insufficient learning of small objects by the algorithm, a key contributor to the low efficiency of small object detection. To address these issues, a feature extraction layer is constructed, consisting of a sequentially connected spatial depth-wise transformation convolutional layer (SPD-Conv) and a feature extraction enhancement module.

[0032] like Figure 3 As shown in the figure, the spatial depth conversion convolution layer is an improved convolution operation designed for small target detection. It aims to solve the problem of small target information loss caused by stride convolution and pooling operations in traditional convolutional neural networks. Its core idea is to achieve downsampling without information loss through spatial segmentation and channel reorganization, so as to better preserve the fine-grained features of small targets. If the size of the image X input to the spatial depth conversion convolution layer is W×W×D1 and the segmentation step is P, then the image X is cut into a series of sub-feature maps at each step. f x,y , which is expressed as: All sub-feature maps f x,y Both are characterized by X ( x,y ) x + i , y+j It can be divided proportionally. Therefore, each sub-feature map can be downsampled by a certain proportion of the feature map X.P interval points, and finally the feature map is evenly divided into P 2 sub-feature maps. For example, when P=2, an 8×8×64 feature map will be divided into 4 4×4×64 sub-feature maps, each sub-map corresponding to a different spatial area of ​​the original image. After completing the spatial segmentation, these sub-feature maps will be spliced ​​and merged in the channel dimension. This step integrates the local feature information originally scattered in different spatial positions into the channel dimension to form an intermediate feature map. At this time, the spatial size of the feature map is reduced to 1 / P of the original, but the number of channels will be increased by P. 2 Taking P=2 as an example, after merging four 4×4×64 sub-images, a new feature map of 4×4×256 is obtained, where 256 is the sum of the original number of channels (64) and the number of divisions (2). 2 =4. Finally, the merged feature map undergoes channel compression through a non-strided convolution operation. The intermediate feature map is processed using a convolution kernel with a stride of 1. While maintaining the spatial resolution, the expanded number of channels is compressed to a more reasonable dimension. For example, 256 channels are compressed to 128 channels through a 1x1 convolution, ultimately outputting a 4×4×128 feature map. This step not only reduces feature dimensionality but also preserves discriminative feature information through learnable convolution parameters. Compared to traditional downsampling methods, the spatial depth conversion convolution layer avoids directly discarding pixel information. Traditional strided convolution or pooling operations skip sampling certain regions, resulting in the neglect of local features of small objects. The spatial depth conversion convolution layer, on the other hand, retains all subregion information through spatial segmentation and then encodes this information into the feature map through channel reorganization. This allows for better extraction of the characteristic details of small face objects.

[0033] like Figure 4 As shown in the figure, the feature extraction enhancement module uses a convolution block (including a convolution layer, a batch normalization layer, and a SiLU activation function) to process the feature map output by the spatial depth conversion convolution layer to obtain an intermediate feature map; the intermediate feature map is split into two feature maps, one part of the feature map is processed in turn by multiple bottleneck modules, and the feature map output by each stacking is spliced ​​with the other part of the feature map in the channel dimension; the splicing result is finally transformed through the convolution block to obtain the feature map output by the feature extraction enhancement module.

[0034] like Figure 5 As shown in the figure, the bottleneck module consists of two sequentially connected convolutional blocks and a detail feature enhancement module. The feature map input to the bottleneck module is processed by the two convolutional blocks and the detail feature enhancement module, and the processed feature map is concatenated with the feature map input to the bottleneck module to obtain the feature map output by the bottleneck module. The two convolutional blocks have the same structure, consisting of a sequentially connected convolutional layer, a batch normalization layer, and a SiLU activation function.

[0035] like Figure 6 As shown, the detail feature enhancement module transforms the input feature map into Decomposed into low-frequency components y L and high frequency component y H , and the high frequency component y H Further divided into horizontal high-frequency features y HL , vertical high-frequency feature y LH and diagonal high-frequency features y HH . The horizontal high-frequency feature y HL , vertical high-frequency feature y LH and diagonal high-frequency features y HH Perform splicing operations and integrate them into overall high-frequency features , which is expressed as: Overall high-frequency characteristics After a convolution block (including convolution, batch normalization and SiLU activation function), the fusion and enhancement are carried out to extract important edge, texture and other detail information. Subsequently, the low-frequency components are combined with the enhanced overall high-frequency features through inverse wavelet transform (IWT) to reconstruct the enhanced feature map after detail enhancement. , which is expressed as: The enhanced feature map Feature map of the input detail feature enhancement module Perform element-by-element addition and fusion to obtain a fusion feature map with rich detail features : If the feature map With enhanced feature map If the sizes are inconsistent, the feature map is expanded , so that it is combined with the enhanced feature map After detail enhancement, the fusion feature map is processed by channel attention mechanism and spatial attention mechanism respectively. Process and get the channel weight and spatial weights . The fusion feature map First, with the channel weight Multiply by the spatial weight Multiply to get the output feature : Channel attention mechanism for fusion feature maps Perform adaptive average pooling to extract a global description of the channel direction avg ; Describe the global avg After two small convolutional networks, the channel weight ω is calculated c : Among them, σ represents the Sigmoid activation function.

[0036] Spatial attention mechanism for fusion feature maps z Perform a 7×7 convolution operation to generate spatial weight ω s : The detail feature enhancement module can effectively enhance the expression of detail features such as texture and edges while preserving spatial structural information, while also improving the representation capabilities of key areas in the feature map. The detail feature enhancement module extracts high-frequency information from the feature map through wavelet transform, including detail texture features in the horizontal, vertical, and diagonal directions, and combines channel attention and spatial attention mechanisms to enhance the expression capabilities of these local detail features. This design enables intermediate features to significantly enhance the expression of details such as texture and edges while preserving the original spatial structural information. This improvement not only improves feature fluidity and fusion efficiency, but also introduces a wavelet detail enhancement mechanism, enabling the feature extraction enhancement module to perform better in fine-grained feature extraction.

[0037] Step 3: Use the dataset to train the face recognition model, and use the trained face recognition model to perform face recognition on the tested image. The recognition results are as follows: Figure 7 shown.

[0038] Step 4: Comparative experiments were conducted on the SCUT_HEAD open source face dataset using the present invention and the Yolov8 model, and the face recognition results are shown in Table 1.

[0039] Table 1 Comparison of results of the present invention and other methods As can be seen from Table 1, the present invention improves the face recognition rate, reduces the probability of false detection and missed detection of small target faces, and improves the recall rate through the feature extraction enhancement module and the newly added 160×160 target detection layer.

[0040] The above-mentioned specific technical processes are used to explain the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A face recognition method based on dynamic weighted wavelet attention, comprising: Construct a dataset containing facial images and preprocess the facial images in the dataset; Build a face recognition model; The face recognition model includes a backbone network, a neck network, and a detection head; the backbone network is used to extract multi-level features from the input image and input them into the neck network for fusion; the detection head is used to convert the fused features into recognition results; and its characteristics are: The backbone network introduces a detail feature enhancement module to extract detail features; in the detail feature enhancement module, the input feature map Decompose into low-frequency components and high-frequency components, and divide the high-frequency components into horizontal high-frequency features, vertical high-frequency features and diagonal high-frequency features; splice the horizontal high-frequency features, vertical high-frequency features and diagonal high-frequency features to obtain the overall high-frequency features; enhance the overall high-frequency features through the convolution block, and reconstruct the low-frequency components and the enhanced high-frequency features to obtain the enhanced feature map; compare the enhanced feature map with the feature map of the input detail feature enhancement module Perform element-by-element addition and fusion to obtain a fused feature map; process the fused feature map through the channel attention mechanism and the spatial attention mechanism respectively to obtain the channel weight and spatial weight; multiply the fused feature map with the channel weight and the spatial weight successively to obtain the feature map output by the feature extraction enhancement module; perform face recognition on the tested image by introducing the face recognition model of the detail feature enhancement module.

2. The face recognition method based on dynamic weighted wavelet attention according to claim 1, characterized in that: The channel attention mechanism performs adaptive average pooling on the fused feature map to extract a global description of the channel direction; the global description is processed by two convolutional networks to obtain channel weights.

3. The face recognition method based on dynamic weighted wavelet attention according to claim 1, characterized in that: The spatial attention mechanism performs a convolution operation on the fused feature map to obtain spatial weights.

4. The face recognition method based on dynamic weighted wavelet attention according to claim 1, characterized in that: The backbone network includes multiple feature extraction layers and a spatial pyramid pooling fast layer connected in sequence. The input image is input into the feature extraction layer after convolution processing for feature extraction. The feature map processed by the multiple feature extraction layers is input into the spatial pyramid pooling fast layer. The feature maps output by the spatial pyramid pooling fast layer and each feature extraction layer are used as the feature maps output by the backbone network.

5. The face recognition method based on dynamic weighted wavelet attention according to claim 4, characterized in that: The feature extraction layer includes a spatial depth conversion convolution layer and a feature extraction enhancement module connected in sequence; the spatial depth conversion convolution layer is used to reorganize the spatial data of the input feature map into channel data, and perform channel compression on the reorganized result through a non-step convolution operation; The feature extraction enhancement module processes the feature map output by the spatial depth conversion convolution layer through a convolution block to obtain an intermediate feature map; splits the intermediate feature map into two feature maps, one of which is processed in turn by multiple bottleneck modules, and the feature map output by each bottleneck module is spliced ​​with the other feature map in the channel dimension; the spliced ​​result is finally transformed through a convolution block to obtain a feature map output by the feature extraction enhancement module.

6. The face recognition method based on dynamic weighted wavelet attention according to claim 5, characterized in that: The bottleneck module includes two convolution blocks connected in sequence and a detail feature enhancement module; the feature map output by the detail feature enhancement module is concatenated with the feature map input to the bottleneck module as the feature map output by the bottleneck module; the two convolution blocks have the same structure, both including a convolution layer, a batch normalization layer and an activation function connected in sequence.

7. The face recognition method based on dynamic weighted wavelet attention according to claim 1, characterized in that: The neck network includes a feature pyramid network and a path aggregation network; The feature pyramid network includes multiple semantic feature fusion layers connected in sequence; the path aggregation network includes multiple detail feature fusion layers connected in sequence; The number of semantic feature fusion layers and detail feature fusion layers is the same.

8. The face recognition method based on dynamic weighted wavelet attention according to claim 7, characterized in that: In the feature pyramid network, each semantic feature fusion layer processes the input feature map through upsampling, and adds the processed feature map to the feature map output by the corresponding feature extraction layer element by element. After processing by the C2f module, the feature map output by the semantic feature fusion layer is obtained. Except for the first semantic feature fusion layer, the input feature map of each semantic feature fusion layer is the feature map output by the previous semantic feature fusion layer. The input feature map of the first semantic feature fusion layer is the feature map output by the spatial pyramid pooling fast layer; In the path aggregation network, except for the last detail feature fusion layer, each detail feature fusion layer processes the input feature map through convolution operation, and adds the processed feature map to the feature map output by the corresponding semantic feature extraction layer element by element. After processing by the C2f module, the feature map output by the detail feature fusion layer is obtained; after convolution processing, the last detail feature fusion layer fuses the feature map output by the spatial pyramid pooling fast layer; except for the first detail feature fusion layer, the feature map input to each detail feature fusion layer is the feature map output by the previous detail feature fusion layer; the feature map input to the first detail feature fusion layer is the feature map output by the last semantic feature extraction layer.

9. The face recognition method based on dynamic weighted wavelet attention according to claim 1, characterized in that: The preprocessing process is: using the mosaic data enhancement method to process the face images in the data set, and performing random scaling, cropping and arrangement.

10. A face recognition system based on dynamic weighted wavelet attention, characterized by: Used to execute the face recognition method based on dynamic weighted wavelet attention according to claim 1; The face recognition system includes an image acquisition module, a preprocessing module and a face recognition module; The image acquisition module is used to collect image data containing human faces and input them into the preprocessing module for processing; The face recognition module is used to identify faces in images.

Images