Human face detection method based on Mamba-YOLO-Face model

Abstract

The invention discloses a face detection method based on a Mamb-YOLO-Face model, deep semantic features are efficiently extracted through a BidGSUSSM backbone network, key information such as textures and edges of a face is captured, background noise interference is effectively reduced, the detection precision is improved, and meanwhile, the network reduces redundant calculation and optimizes the structure, so that the detection accuracy is improved. Compared with the prior art, the C2f-BidGSUSSM neck network has the advantages that model parameters and calculation complexity are reduced, balance of precision and efficiency is realized, in addition, the BidGSUSSM adaptively adjusts sensitivity to illumination and postures by using multi-scale feature extraction and an attention mechanism, false detection / missing detection caused by illumination changes or side faces is reduced, and the C2f-BidGSUSSM neck network has the advantages that the C2f-BidGSUSSM neck network is optimized through cross-scale feature fusion, so that the accuracy and accuracy of the C2f-BidGSUSSM neck network is improved. Compared with the prior art, the 5Point Landmark Regression Head network has the advantages that the detection capability on small targets and shielded faces is enhanced, the omission ratio is remarkably reduced, the 5Point Landmark Regression Head network optimizes the detection frame positioning precision by regressing face key points, and the robustness on posture change and shielded faces is improved by using key point constraints.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of face detection, in particular to a face detection method based on a Mamba-YOLO-Face model. BACKGROUND

[0002] With the rapid development of computer vision technology, face detection, as one of its core tasks, has been widely applied in the fields of security monitoring, human-computer interaction, identity authentication and the like. However, the traditional face detection method often has low detection accuracy and poor real-time performance when facing complex scenes (such as light changes, occlusions, small targets and the like). In recent years, the rise of deep learning technology has provided a new solution for face detection. By constructing a deep convolutional neural network model, deep features in an image can be automatically learned, thereby significantly improving the performance of face detection. The YOLO series model, as a classic algorithm in the field of target detection, has attracted widespread attention due to its efficient real-time performance and good detection accuracy. The YOLO model realizes end-to-end training and prediction by converting the target detection task into a regression problem, thereby greatly improving the detection speed. However, the traditional YOLO model still has certain limitations when processing small targets, dense targets and complex backgrounds. SUMMARY

[0003] Therefore, the application provides a face detection method based on a Mamba-YOLO-Face model, which can effectively solve the defects of the prior art in processing small targets, dense targets and complex backgrounds.

[0004] The technical scheme of the application is as follows: A face detection method based on a Mamba-YOLO-Face model, specifically comprising the following steps: Constructing a Mamba-YOLO-Face model, wherein the Mamba-YOLO-Face model comprises a backbone network, a neck network and a head network, the backbone network adopts a BidGSUSSM backbone network, the neck network adopts a C2f-BidGSUSSM neck network, and the head network adopts a 5Point Landmark Regression Head head network; Obtaining image data to be recognized from an image acquisition device or an image storage device; Preprocessing the obtained image data; Inputting the preprocessed image data into the Mamba-YOLO-Face model, and outputting a face detection result by the Mamba-YOLO-Face model.

[0005] As a further optional solution of the face detection method based on the Mamba-YOLO-Face model, the BidGSUSSM stem network comprises: a Stem module, configured to perform down-sampling operation on an input image to generate a two-dimensional feature map with a resolution of H / 4 and W / 4; a GSUSSM module, configured to perform image feature extraction on the two-dimensional feature map after the Stem module; a first CBS module, configured to perform down-sampling and channel number adjustment on the image feature after the GSUSSM module; a BidGSUSSM module, configured to perform forward and reverse processing on the image feature that has been down-sampled and channel number adjusted by the first CBS module, and fuse the features processed in the forward and reverse directions by using a GSU module to achieve further lightweight extraction of the features and obtain an initial feature map.

[0006] As a further optional solution of the face detection method based on the Mamba-YOLO-Face model, the C2f-BidGSUSSM neck network comprises: a second CBS module, configured to perform convolution, batch normalization and activation function processing on the initial feature map input to the neck network, adjust the channel number of the initial feature map, and perform preliminary feature extraction; a feature segmentation layer, configured to divide the preliminary feature of the initial feature map processed by the second CBS module into two along the channel dimension to generate two feature branches, one of which directly enters a feature concatenation layer, and the other of which enters a C2f-BidGSUSSM module group for deep-level feature extraction; the C2f-BidGSUSSM module group is composed of one or more C2f-BidGSUSSM modules, and each C2f-BidGSUSSM module is configured to perform forward and reverse processing on the input feature and fuse the bidirectional features by using a GSU module to achieve further lightweight extraction of the features; a feature concatenation layer, configured to concatenate the feature branches processed by the C2f-BidGSUSSM module group with the feature branch that directly enters the feature concatenation layer to fuse feature information at different levels; a channel compression module, configured to compress the channel number of the concatenated feature map to obtain a final feature map.

[0007] As a further optional solution of the face detection method based on the Mamba-YOLO-Face model, the 5Point Landmark Regression Head head network comprises: A global deep pooling layer is used to perform pooling operations on the final feature map by calculating the maximum or average value along the channel dimension, generating a one-dimensional feature vector that retains the global information of the final feature map. The fully connected layer receives the one-dimensional feature vector output from the global deep pooling layer and maps the feature vector to the output space through a linear transformation, outputting the predicted coordinates of five key points.

[0008] As a further alternative to the face detection method based on the Mamba-YOLO-Face model, during the face detection process using the Mamba-YOLO-Face model, the Non-Maximum Suppression (NMS) algorithm is used to remove overlapping detection boxes.

[0009] As a further optional solution to the face detection method based on the Mamba-YOLO-Face model, the Non-Maximum Suppression (NMS) algorithm removes overlapping detection boxes, specifically including: The Mamba-YOLO-Face model generates a set of candidate bounding boxes through the BidGSUSSM backbone network, the C2f-BidGSUSSM neck network, and the 5Point Landmark Regression Head network. Each box contains location information, confidence score, and possible key point information. The candidate detection boxes output by the model are sorted from high to low confidence. The detection box with the highest confidence is selected as the baseline box, and the intersection-union ratio (IUR) of the baseline box with all other boxes is calculated. If the IUR of a box with the baseline box exceeds a preset threshold, it is considered that the box and the baseline box detect the same target, and the box is removed from the candidate box list.

[0010] A face detection system based on the Mamba-YOLO-Face model includes: The building module is used to build the Mamba-YOLO-Face model, which includes a backbone network, a neck network, and a head network. The backbone network adopts the BidGSUSSM backbone network, the neck network adopts the C2f-BidGSUSSM neck network, and the head network adopts the 5Point Landmark Regression Head network. The acquisition module is used to acquire image data to be identified from an image acquisition device or an image storage device; The preprocessing module is used to preprocess the acquired image data; The detection module is used to input the preprocessed image data into the Mamba-YOLO-Face model, and the Mamba-YOLO-Face model outputs the face detection results.

[0011] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of any of the above-described face detection methods based on the Mamba-YOLO-Face model.

[0012] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described face detection methods based on the Mamba-YOLO-Face model.

[0013] The beneficial effects of this invention are as follows: The BidGSUSSM backbone network can extract deep semantic features from images more efficiently, especially excelling at capturing key information such as facial texture and edges, reducing background noise interference, and thus improving detection accuracy. Simultaneously, by reducing redundant computation and optimizing the network structure, the BidGSUSSM backbone network reduces the number of model parameters and computational complexity while maintaining accuracy. Furthermore, through multi-scale feature extraction and attention mechanisms, the BidGSUSSM backbone network can adaptively adjust its sensitivity to different lighting conditions and facial poses, reducing false positives / false negatives caused by lighting changes or side profiles. The C2f-BidGSUSSM neck network, through cross-scale feature fusion, enables the model to simultaneously perceive facial features at different scales, significantly enhancing the detection capabilities for small targets and occluded faces, and reducing the false negative rate. The 5Point LandmarkRegression Head network, by regressing facial key points, not only optimizes the localization accuracy of the detection box but also improves robustness to pose changes and occluded faces through key point constraints. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a face detection method based on the Mamba-YOLO-Face model according to the present invention. Figure 2This is a schematic diagram of the composition of a face detection system based on the Mamba-YOLO-Face model according to the present invention; Figure 3 This is a schematic diagram of the composition of a computing device according to the present invention; Figure 4 This is a schematic diagram of the Mamba-YOLO-Face model. Figure 5 This is a schematic diagram of the BidGSUSSM backbone network. Figure 6 This is a schematic diagram of the Stem module; Figure 7 This is a schematic diagram of the structure of the first CBS module; Figure 8 This is a schematic diagram of the C2f-BidGSUSSM neck network. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0017] refer to Figures 1 to 8 A face detection method based on the Mamba-YOLO-Face model, specifically including: A Mamba-YOLO-Face model is constructed, which includes a backbone network, a neck network, and a head network. The backbone network adopts the BidGSUSSM backbone network, the neck network adopts the C2f-BidGSUSSM neck network, and the head network adopts the 5Point Landmark Regression Head network. Obtain the image data to be identified from the image acquisition device or image storage device; The acquired image data is preprocessed (e.g., resizing, normalizing, etc.) to adapt to the input requirements of the Mamba-YOLO-Face model; The preprocessed image data is input into the Mamba-YOLO-Face model, which outputs the face detection results.

[0018] In this embodiment, the BidGSUSSM backbone network can extract deep semantic features from images more efficiently, especially adept at capturing key information such as facial texture and edges, reducing background noise interference and thus improving detection accuracy. Simultaneously, by reducing redundant computation and optimizing the network structure, the BidGSUSSM backbone network reduces the number of model parameters and computational complexity while maintaining accuracy. Furthermore, through multi-scale feature extraction and attention mechanisms, the BidGSUSSM backbone network can adaptively adjust its sensitivity to different lighting conditions and facial poses, reducing false positives / false negatives caused by lighting changes or side profiles. The C2f-BidGSUSSM neck network, through cross-scale feature fusion, enables the model to simultaneously perceive facial features at different scales, significantly enhancing the detection capabilities for small targets and occluded faces, and reducing the false negative rate. The 5Point Landmark RegressionHead head network, by regressing facial key points, not only optimizes the localization accuracy of the detection box but also improves robustness to pose changes and occluded faces through key point constraints.

[0019] It should be noted that the BidGSUSSM backbone network is adopted. This network downsamples through the Stem module to generate a two-dimensional feature map with a specific resolution, which then enters the GSUSSM module and is further downsampled through the CBS module. The BidGSUSSM module, as the core module, is used to capture a richer gradient information flow and reduce computational complexity. The C2f-BidGSUSSM neck network is adopted. Based on the original C2f, the Bottleneck module is replaced with the BidGSUSSM module. The feature learning ability is enhanced by channel segmentation and the BidGSUSSM module, and the original information is preserved by residual connection. A 5Point Landmark Regression Head network is adopted, which adds the ability to predict the location of 5 key points on the face (including the two eyes, the tip of the nose and the two corners of the mouth) on the basis of the original YOLOv8 head network, so as to improve the accuracy of face alignment and expression recognition. In addition, the BidGSUSSM backbone network is responsible for extracting basic features from the input image. It extracts multi-scale features of the image step by step through a series of convolutional layers, normalization layers and activation functions, as well as the core BidGSUSSM module. The extracted feature maps are then passed to the C2f-BidGSUSSM neck network. These feature maps contain low- to high-level features of the image, providing rich information for further processing by the neck network. The C2f-BidGSUSSM neck network receives feature maps from the backbone network and performs feature fusion and enhancement through its unique C2f-BidGSUSSM structure. This structure combines the feature segmentation and fusion capabilities of the C2f module with the bidirectional feature extraction capabilities of the BidGSUSSM module, enabling the neck network to more effectively utilize the features extracted by the backbone network. By stacking multiple C2f-BidGSUSSM units, the neck network forms a multi-level feature fusion network, further extracting and fusing features at different scales, providing more discriminative feature representations for the head network. After feature fusion and enhancement by the neck network, the resulting feature maps contain rich image information and contextual relationships. These feature maps are then passed to the 5Point Landmark Regression Head network. The head network receives these feature maps and prepares to perform the final face detection and keypoint prediction tasks. Based on the received feature maps, the 5Point Landmark Regression Head network outputs the face detection results (including location and confidence) and the predicted locations of five keypoints (two eyes, nose tip, and two corners of the mouth) through independent convolutional layers and prediction layers. The design of the head network decouples the classification task, bounding box regression task, and keypoint prediction task, thereby accelerating network convergence and improving detection accuracy.

[0020] Preferably, the BidGSUSSM backbone network includes: The Stem module is used to downsample the input image. This module typically contains components such as convolutional layers, batch normalization layers, and activation functions. Through the combination of these components, the input image is initially extracted and its size reduced, generating a two-dimensional feature map with a resolution of H / 4 and W / 4, where H and W represent the height and width of the input image, respectively. The GSUSSM module, following the Stem module, is used to extract image features from the two-dimensional feature map; The first CBS module, following the GSUSSM module, is used to downsample and adjust the number of channels of the image features processed by the GSUSSM module. The first CBS module contains convolutional layers, batch normalization layers, and activation functions. Downsampling is achieved by adjusting the size, stride, and padding of the convolutional kernels, while the number of channels of the feature map is changed by adjusting the number of convolutional kernels. The BidGSUSSM module is used to perform forward and reverse processing on image features that have been downsampled and adjusted by the first CBS module. By introducing a bidirectional processing mechanism, the BidGSUSSM module can capture a richer gradient information stream. At the same time, the module uses the GSU (Global State Update) module to fuse the features processed by forward and reverse processing to achieve further lightweight feature extraction and obtain the initial feature map.

[0021] In this embodiment, the Stem module reduces the input image resolution from H×W to H / 4×W / 4 through convolution and pooling operations, significantly reducing subsequent computation (computation is proportional to the square of the resolution) while preserving basic semantic information (such as edges and textures). It also expands the number of channels to a suitable range (e.g., 64 channels) through convolution kernels, providing a multi-dimensional representation space for subsequent feature extraction and avoiding the computational explosion caused by directly processing high-resolution images. The GSUSSM module, through grouped convolution and spatial attention mechanisms, can simultaneously capture local details (such as facial organ contours) and global structure (such as the overall facial layout), enhancing the robustness of features to pose and occlusion. This is achieved through residual connections... The Connection module alleviates the vanishing gradient problem, enabling deep networks to reuse shallow features, thus improving training stability and convergence speed. The first CBS module further reduces the feature map resolution (e.g., H / 8×W / 8) through stride convolution, reducing the computational complexity of subsequent layers while expanding the receptive field to capture a wider range of contextual information. It dynamically adjusts the number of channels (e.g., from 64 to 128) through 1×1 convolution, balancing feature dimensionality and computational cost, and avoiding redundant computation due to excessive channel count. The BidGSUSSM module, through forward and backward processing (e.g., bidirectional LSTM or self-attention mechanism), allows the model to simultaneously learn the temporal dependencies of features (e.g., spatial relationships of facial key points) and hierarchical structures (e.g., abstraction from local to global), enhancing feature representation capabilities. It dynamically weights and fuses bidirectional features through the Global Feature Selection Unit (GSU), filtering redundant information and strengthening key features, achieving feature lightweighting (e.g., reducing the number of parameters by 30%) while improving detection accuracy. It reduces cross-channel communication overhead through grouped convolution and channel shuffle, and further reduces computational cost by combining depthwise separable convolution.

[0022] It should be noted that the working principle of the GSU module is as follows: Feature enhancement: ; in, Indicates the first l In-layer tasks k Features and These are the weight matrix and the bias vector, respectively, and σ is the activation function (such as ReLU, Sigmoid, etc.). Through this linear transformation and non-linear activation, GSU ​​can transform the input features and enhance their expressive power. This transformation enables the features to better adapt to the needs of different tasks and extract more discriminative information. Task Collaboration: ; in, It is the first l Tasks in +1 level j The characteristic of C is that C is the set of tasks. These are the weights calculated earlier, and ⊙ represents element-wise multiplication; This formula indicates that the task j In the l The features of layer +1 not only depend on its own first... l Layer features It also depends on other tasks. k (k ≠ j) of the first l Layer features Through weight Perform weighted fusion; This fusion mechanism enables different tasks to share information and leverage the features of other tasks to enhance the feature representation of their own tasks, thereby improving the collaborative effect between tasks.

[0023] Preferably, the C2f-BidGSUSSM neck network includes: The second CBS module is used to perform convolution, batch normalization, and activation function processing on the initial feature map input to the neck network, adjust the number of channels in the initial feature map, and perform preliminary feature extraction. The feature segmentation layer is used to divide the initial features of the initial feature map processed by the second CBS module into two along the channel dimension, generating two feature branches. One feature branch directly enters the feature splicing layer, and the other feature branch enters the C2f-BidGSUSSM module group for deep feature extraction. The C2f-BidGSUSSM module group consists of one or more C2f-BidGSUSSM modules. Each C2f-BidGSUSSM module is used to process the input features in both forward and reverse directions, and uses the GSU module to fuse the bidirectional features to achieve further lightweight feature extraction. The feature concatenation layer is used to concatenate the feature branches processed by the C2f-BidGSUSSM module group with the directly entered feature branches, thus fusing feature information from different levels. The channel compression module is used to compress the number of channels in the stitched feature map to obtain the final feature map.

[0024] In this embodiment, the second CBS module dynamically adjusts the number of channels in the input feature map through convolutional operations (e.g., compressing from 256 channels to 128 channels), reducing the computational load of subsequent modules while preserving key semantic information. Batch normalization accelerates model convergence and improves training stability. Activation functions (e.g., SiLU) introduce non-linearity to enhance feature representation. Lightweight convolutional layers extract shallow features (e.g., edges, textures), providing a foundation for subsequent deep feature extraction. The feature segmentation layer divides the feature map along the channel dimension, generating two independent branches to process shallow and deep features respectively, avoiding information loss due to feature mixing. One branch directly enters the feature concatenation layer, preserving the original features, while the other branch enters the C2f-BidGSUSSM module group for deep processing, achieving feature decoupling and differential extraction. Each C2f-BidGSUSSM module captures the temporal dependence and spatial relationship of features through forward and reverse processing (e.g., combining self-attention mechanisms and recurrent neural networks), enhancing the robustness of features to pose and occlusion. The Global Feature Selection Unit (GSU) further enhances this. The model dynamically weights and fuses bidirectional features, filtering redundant information and strengthening key features. This achieves feature lightweighting (e.g., reducing the number of parameters by 30%) while improving detection accuracy. By stacking multiple C2f-BidGSUSSM modules, the model can gradually extract deeper semantic information, adapting to detection tasks of varying complexity. The feature concatenation layer concatenates shallow feature branches that directly enter the model with feature branches that have undergone deep processing, fusing low-level features (e.g., edges) and high-level features (e.g., semantic information) to improve feature comprehensiveness. The concatenation operation maintains the continuity of gradient propagation, avoiding training instability caused by feature loss. The channel compression module compresses the concatenated high-dimensional feature map to a suitable range (e.g., reducing it from 256 channels to 128 channels) through 1×1 convolutions, reducing the computational load of subsequent head networks. While compressing the number of channels, the model adjusts the feature distribution through convolution kernel weights, strengthening key channel information and suppressing noise.

[0025] Preferably, the 5Point Landmark Regression Head network includes: A global deep pooling layer is used to perform pooling operations on the final feature map by calculating the maximum or average value along the channel dimension, generating a one-dimensional feature vector that retains the global information of the final feature map. The fully connected layer receives the one-dimensional feature vector output from the global deep pooling layer and maps the feature vector to the output space through a linear transformation, outputting the predicted coordinates of five key points.

[0026] In this embodiment, the final feature map (e.g., H×W×C) is compressed into a one-dimensional feature vector (e.g., 1×C) by calculating the maximum or average value along the channel dimension, preserving the global spatial distribution and semantic information of the feature map. The pooling operation compresses only along the spatial dimensions (H, W), while the number of channels (C) remains unchanged, ensuring that the subsequent fully connected layers can fully utilize the multi-channel features. Global pooling reduces the impact of local noise (e.g., background interference, occlusion) on keypoint prediction by aggregating global information, thus improving the robustness of the model. The fully connected layer linearly transforms the one-dimensional feature vector (1×C) into the predicted coordinates of five keypoints (e.g., 1×10, each keypoint containing x and y coordinates) through the weight matrix and bias term. The weight matrix of the fully connected layer can be regarded as a mapping template from the feature vector to the keypoint coordinates. By training and optimizing the template parameters, the model can adapt to different facial poses and expressions. Compared with convolutional layers, fully connected layers have fewer parameters in keypoint regression tasks (e.g., only 10×C parameters are needed), resulting in higher computational efficiency.

[0027] Preferably, during face detection using the Mamba-YOLO-Face model, the Non-Maximum Suppression (NMS) algorithm is used to remove overlapping detection boxes.

[0028] Preferably, the NMS (Non-Maximum Suppression) algorithm for removing overlapping detection boxes specifically includes: The Mamba-YOLO-Face model generates a set of candidate bounding boxes through the BidGSUSSM backbone network, the C2f-BidGSUSSM neck network, and the 5Point Landmark Regression Head network. Each box contains location information (such as the coordinates of the top left and bottom right corners), confidence (the probability that the model predicts that the box contains the target), and possible key point information (in face detection tasks, this includes the coordinates of key points such as the eyes, the tip of the nose, and the corners of the mouth). The candidate detection boxes output by the model are sorted from high to low confidence. The purpose of sorting is to prioritize the detection boxes with high confidence, because they are more likely to correspond to the real target. Select the detection box with the highest confidence as the reference box, and calculate the intersection-union ratio (IoU) of the reference box with all other boxes. IoU is an indicator that measures the degree of overlap between two boxes. The calculation formula is the intersection area of ​​the two boxes divided by the union area. If the intersection-union ratio (IUU) of a bounding box with the reference bounding box exceeds a preset threshold, it is considered that the bounding box and the reference bounding box are detecting the same target. In this case, the bounding box is removed from the candidate box list to avoid duplicate detection. Repeat the above steps until all candidate boxes have been processed. The final remaining detection boxes are the results of the NMS algorithm. They correspond to different targets and do not overlap or have very low overlap with each other.

[0029] In this embodiment, the optimal bounding box for each target is selected from a set of candidate detection boxes generated by the model, and redundant or duplicate boxes are removed to ensure the uniqueness and accuracy of the detection results. The model may generate multiple high-confidence boxes for the same target (e.g., at different scales or locations). NMS uses an IoU threshold to filter, ensuring that only one optimal bounding box is retained for each target, improving the simplicity of the detection results, reducing post-processing complexity, and avoiding misjudgments caused by redundant boxes in multi-face scenarios (e.g., identifying the same person as multiple people). NMS only needs to calculate IoU and sort, with a complexity of O(n²) (where n is the number of candidate boxes), which is more efficient than other post-processing methods (e.g., bounding box merging based on graph models). In resource-constrained devices (such as embedded cameras), NMS's fast processing capability ensures real-time detection. When NMS filters redundant boxes through IoU, it implicitly requires that the retained boxes be spatially close to the real target, thereby indirectly improving the accuracy of keypoint prediction (e.g., the coordinates of eyes and nose are closer to their real positions). If keypoint detection and object detection are trained together, NMS can enhance the consistency between the two and reduce keypoint drift caused by box position deviations. In partially occluded scenarios, NMS filters overlapping boxes through IoU thresholding to avoid box splitting caused by occlusion (e.g., misdetecting a face as two separate boxes). NMS is not sensitive to target scale and is suitable for face detection of different sizes.

[0030] A face detection system based on the Mamba-YOLO-Face model includes: The building module is used to build the Mamba-YOLO-Face model, which includes a backbone network, a neck network, and a head network. The backbone network adopts the BidGSUSSM backbone network, the neck network adopts the C2f-BidGSUSSM neck network, and the head network adopts the 5Point Landmark Regression Head network. The acquisition module is used to acquire image data to be identified from an image acquisition device or an image storage device; The preprocessing module is used to preprocess the acquired image data; The detection module is used to input the preprocessed image data into the Mamba-YOLO-Face model, and the Mamba-YOLO-Face model outputs the face detection results.

[0031] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of any of the above-described face detection methods based on the Mamba-YOLO-Face model.

[0032] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described face detection methods based on the Mamba-YOLO-Face model.

[0033] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A face detection method based on the Mamba-YOLO-Face model, characterized in that, Specifically, it includes: A Mamba-YOLO-Face model is constructed, which includes a backbone network, a neck network, and a head network. The backbone network adopts the BidGSUSSM backbone network, the neck network adopts the C2f-BidGSUSSM neck network, and the head network adopts the 5Point Landmark Regression Head network. Obtain the image data to be identified from the image acquisition device or image storage device; Preprocess the acquired image data; The preprocessed image data is input into the Mamba-YOLO-Face model, which outputs the face detection results.

2. The face detection method based on the Mamba-YOLO-Face model according to claim 1, characterized in that, The BidGSUSSM backbone network includes: The Stem module is used to downsample the input image and generate a two-dimensional feature map with a resolution of H / 4 and W / 4. The GSUSSM module, following the Stem module, is used to extract image features from the two-dimensional feature map; The first CBS module, following the GSUSSM module, is used to downsample image features and adjust the number of channels; The BidGSUSSM module is used to perform forward and reverse processing on the image features downsampled and adjusted by the first CBS module, and then use the GSU module to fuse the forward and reverse processed features to achieve further lightweight feature extraction and obtain the initial feature map.

3. The face detection method based on the Mamba-YOLO-Face model according to claim 2, characterized in that, The C2f-BidGSUSSM neck network includes: The second CBS module is used to perform convolution, batch normalization, and activation function processing on the initial feature map input to the neck network, adjust the number of channels in the initial feature map, and perform preliminary feature extraction. The feature segmentation layer is used to divide the initial features of the initial feature map processed by the second CBS module into two along the channel dimension, generating two feature branches. One feature branch directly enters the feature splicing layer, and the other feature branch enters the C2f-BidGSUSSM module group for deep feature extraction. The C2f-BidGSUSSM module group consists of one or more C2f-BidGSUSSM modules. Each C2f-BidGSUSSM module is used to process the input features in both forward and reverse directions, and uses the GSU module to fuse the bidirectional features to achieve further lightweight feature extraction. The feature concatenation layer is used to concatenate the feature branches processed by the C2f-BidGSUSSM module group with the directly entered feature branches, thus fusing feature information from different levels. The channel compression module is used to compress the number of channels in the stitched feature map to obtain the final feature map.

4. The face detection method based on the Mamba-YOLO-Face model according to claim 3, characterized in that, The 5Point Landmark Regression Head network includes: A global deep pooling layer is used to perform pooling operations on the final feature map by calculating the maximum or average value along the channel dimension, generating a one-dimensional feature vector that retains the global information of the final feature map. The fully connected layer receives the one-dimensional feature vector output from the global deep pooling layer and maps the feature vector to the output space through a linear transformation, outputting the predicted coordinates of five key points.

5. The face detection method based on the Mamba-YOLO-Face model according to claim 4, characterized in that, During face detection using the Mamba-YOLO-Face model, the Non-Maximum Suppression (NMS) algorithm is used to remove overlapping detection boxes.

6. The face detection method based on the Mamba-YOLO-Face model according to claim 5, characterized in that, The Non-Maximum Suppression (NMS) algorithm removes overlapping detection boxes, specifically including: The Mamba-YOLO-Face model generates a set of candidate bounding boxes through the BidGSUSSM backbone network, the C2f-BidGSUSSM neck network, and the 5PointLandmark Regression Head network. Each box contains location information, confidence score, and possible key point information. The candidate detection boxes output by the model are sorted from high to low confidence. The detection box with the highest confidence is selected as the baseline box, and the intersection-union ratio (IUR) of the baseline box with all other boxes is calculated. If the IUR of a box with the baseline box exceeds a preset threshold, it is considered that the box and the baseline box detect the same target, and the box is removed from the candidate box list.

7. A face detection system based on the Mamba-YOLO-Face model, characterized in that, include: The building module is used to build the Mamba-YOLO-Face model, which includes a backbone network, a neck network, and a head network. The backbone network adopts the BidGSUSSM backbone network, the neck network adopts the C2f-BidGSUSSM neck network, and the head network adopts the 5Point Landmark Regression Head network. The acquisition module is used to acquire image data to be identified from an image acquisition device or an image storage device; The preprocessing module is used to preprocess the acquired image data; The detection module is used to input the preprocessed image data into the Mamba-YOLO-Face model, and the Mamba-YOLO-Face model outputs the face detection results.

8. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the face detection method based on the Mamba-YOLO-Face model as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the face detection method based on the Mamba-YOLO-Face model as described in any one of claims 1-6.

Images