A method of mask segmentation that combines the most important gender-related characteristics and constraints (2023)

With the global COVID-19 pandemic, masks have become an essential element in public places, and many people wear them on a daily basis to protect themselves and others [1-5]. This is a challenge for access control systems, payment systems and other security and facilities based on facial recognition technology. Traditional facial recognition technology faces many challenges in recognizing faces in masks, such as partially obscured facial features, light reflections caused by masks, etc. [6-11]. To solve this problem, it is necessary to develop a method that can effectively recognize and segment faces in masks [12-20]. By exploring a face mask segmentation method that combines relevant gender-related characteristics and constraints, facial recognition accuracy under mask occlusion conditions can be improved, thus playing an important role in public safety, financial payments, access control systems, AI monitoring, health management etc...

In today's specific conditions, due to the pandemic, people have to wear masks in their daily activities. In some cases, people were forced to remove their masks during the performance. Consider, for example, the traditional approach of facial recognition systems to presence monitoring where individuals are forced to remove their masks, which is not apparent in the current context. Priya et al. [21] have created a system that helps individuals register their presence without having to remove their masks. The model uses the Caffe model for face detection, the CNN model for face recognition, and the Haar Cascade classifier. This model was created for use in real-time applications. The accuracy of the system design is 90%. Rao and so on. [22] proposed FMRS-CFR (Face Mask Recognition System - Centerface Resnet), an epidemic mask recognition system based on the fusion of multiple algorithms to adapt to multi-scenario applications. The work uses central face detection models and the Resnet50 classification. A system that dynamically maintains multiple adjustments to outdoor scenes was constructed, then transplanted into the Atlas 200 development kit, and video quantification of a dozen different scenes was performed. The experimental results show that the FMRS-CFR system can achieve a recognition accuracy rate of 99.88%, which greatly improves the recognition rate without a mask or wearing a mask correctly to some extent, and achieves the goal of effectively helping to prevent and control epidemics. . Kaliapan et al. [23] divided people into three categories, such as wearing a mask, not wearing a mask, and wearing a mask incorrectly. The dataset was tested using three different variants of the object detection model, namely YOLOv4, Tiny YOLOv4 and YOLOv5. Experimental results show that the performance of the YOLOv5 model is better than the other two models, with the highest mAP value reaching 99.40%. Ramachandra and Marcel [24] were the first to explore the use of face-fitting 3D silicone masks as a means to create facial morphing attacks. A systematic study has been presented to measure the potential of (digital) mask deformation attacks on commercial and academic FRSs. To this end, we created a new data set using eight custom-made 3D silicone masks and corresponding real face images captured by three different smartphones. Mask deformation is performed using a landmark-based approach, and the newly constructed dataset includes 635 real, 1034 masks, and 613 masks of deformed face images. Extensive experiments are being conducted to determine the attack potential and detect mask deformation attacks on the FRS.

Based on existing research results, the current face mask segmentation method has improved the face recognition accuracy of mask wearers to some extent. However, there are still some disadvantages and disadvantages. Under different lighting conditions, it may be difficult to extract important features, which may lead to reduced recognition performance. Gender constraints help improve the accuracy of mask segmentation, but in some special cases, such as transgender people or people with ambiguous gender expressions, gender can be misjudged and affect segmentation results. Deep learning methods can add computational complexity, which affects real-time performance. In scenarios where you need to process a large number of images quickly, this method may not meet your real-time requirements. Therefore, this article explores a face mask segmentation method that combines relevant gender-related characteristics and constraints.

To enable the model to realize real-time face detection on the hardware platform, section 2 of this paper optimizes the structure of a multi-purpose cascaded convolutional neural network by introducing a depth-separable convolution and completes the face detection task by combining salient features and gender Section 3 Section 1 completes the extraction of the mask area and gives technical steps for mask extraction based on spectral features. Experimental results confirm the effectiveness of the constructed model.

To achieve real-time face detection on a hardware platform, this paper optimizes a multi-purpose cascaded convolutional neural network architecture by introducing depth-separated convolutions. Depth-separated convolution is a computationally efficient and effective convolution method that separates spatial convolution and channel convolution and still effectively extracts image features while achieving lower performance and computational complexity. This allows the algorithm to achieve higher recognition accuracy at lower computational cost in a face detection task that includes gender characteristics and constraints. Reducing model parameters makes it easier to implement the model on hardware platforms such as embedded devices and mobile devices. This enables the wide application of face detection methods that include gender-related characteristics and constraints on different hardware platforms. The multi-tasking cascade convolutional neural network structure adopted in this paper can handle multiple tasks simultaneously, such as mask detection and gender recognition. This allows the algorithm to implement multi-attribute recognition that combines gender-specific characteristics and constraints in face detection tasks, improving recognition versatility.

The following formulas provide calculations of parameter sizesRice_HisIRice_DIn a standard weave, it is assumed that the length of the side of the weave nucleus is expressed asC_G, the number of input and output channels is expressed asMan_existIMan_{go out}, this is:

$M_t=C_G \times C_G \times D_{i n} \times D_{\text {out }}$ (1)

$M_d=F_{\text {out }} \times Q_{\text {out }} \times D_{\text {out }} \times C_G \times C_G \times D_{\text {in }}$ (2 )

Figure 1 shows the procedure for calculating Mt and Md in a standard weave. However, the formula for the parameter quantRice_Hisand number of calculationsRice_DexistDeep wisdomThe calculation of the weave used, which can be separated by depth, is as follows:

$M_t^{D W}=C_G \times C_G \times D_{i n}$ (3)

$M_d^{D W}=C_G \times C_G \times D_{\text {in }} \times F_{\text {out }} \times Q_{\text {out }}$ (4)

The following formulas provide calculations of parameter sizes and budget sizesPointWise:

$M_t^{P W}=1 \times 1 \times D_{\text {in }} \times D_{\text {out }}$ (5)

$M_d^{P W}=F_{o u t} \times Q_{\text {out }} \times 1 \times 1 \times D_{\text {in }}$ (6)

Add the above formulas one by one to get the total number of parameters and the total number of strands split deep. The calculation procedure is presented in the following formula:

$M_t=M_t^{D W}+M_t^{P W}$ (7)

$M_d=M_d^{D W}+M_d^{P W}$ (8)

As can be seen from the above fabrication process, the depth-separated weave significantly reduces the number of model parameters and calculations based on the traditional standard weave. Suppose that the ratio of the total number of parameters of the deep split weave to the standard weave is expressed asB, the share in the total settlement amount isC, get the following calculation formula:

$\beta=\frac{M^{C Q}+T Q}{M}=\frac{C_G \times C_G \times D_{\text {in }}+D_{\text {in }} \times D_{\文本 {out }}}{C_G \times C_G \times D_{\text {in }} \times D_{\text {out }}}=\frac{C_G^2+D_{\text {out }}}{ C_G^2 \times D_{\text {输出}}}$ (9)

$\gamma=\frac{M^{D W}+T Q}{M}=\frac{C_G^2+D_{\text {out}}}{D_{\text {out}} \times C_G^2} =\frac{1}{D_{\text {output}}}=\frac{1}{C_G^2}$ (10)

Separate training of facial characteristics and gender attributes requires the implementation of a facial recognition network and a gender recognition network. The disadvantage of this approach is that the hardware must read the weights of both networks, which doubles the total volume of parameters compared to a single network. Therefore, this article designed a multi-attribute facial recognition network that integrates iconic gender-specific features and constraintsVGG-11as the backbone of the network. The total number of parameters can be reduced by integrating gender-specific characteristics and constraints into a single network. This can reduce the load on your hardware and improve your computer's performance. The integration of distinctive objects and gender constraints in one network also allows for better use of shared representations of objects in the network. This helps improve the performance of multi-attribute recognition tasks and avoids the redundancy of separate training. Also,VGG-11networks, compared to other deeper convolutional neural networks such asVGG-16,VGG-19etc.), which has low computational complexity and can maintain high recognition accuracy while reducing run time.

In order to further reduce the number of parameters and obtain a smaller input data size, this paper tends to the originalVGG-11network. Figure 2 shows the optimized network structure. by erasingThe nucleus of the plexus 8,FC, IFC2The number of model parameters is significantly reduced, which reduces the load on hardware resources, reduces computational costs and improves computational speed. Resizing the input image to 224×224 reduces computing and memory requirements. Thanks to this, the network can achieve high performance even in conditions of low computational resources. The optimized network structure is more compact, which helps to increase the flexibility of model implementation on different devices. changing the sizeFC3Layers up to 1×1×22, the network can be tailored to multi-attribute recognition tasks. This allows the network to simultaneously process multiple tasks such as gender characteristics and constraints, improving the versatility and efficiency of recognition.

This article attempts to combine the most important features of the face and gender into a unique structure of the output network. The specific representation and combination method are given below: Assume existsSoSignificant features (such as key points, facial expressions, etc.) that can vary from 1 toSoFor gender we can use the same coding as above, 0 for females and N+1 for males. Then add the face function value and gender value to get (So+1)*2 dimensional output space. In the last fully connected layerVGGgrid, the output size is changed to 1×1×((So+1)*2), the probability value is obtained with softmax. In the processing phase, the output can be obtained from (So+1)*2 Probability values ​​and corresponding salient features and gender classification results can be calculated. For example, given the 10 most significant characteristics, gender is coded as 0 (female) and 1 (male), resulting in 22 scores of size 1 × 1 × 22. By processing the output probabilities, a classification result can be obtained that includes both significant characteristics, and gender attributes. The specific calculation method is shown in the formula below. Suppose that the probability that the network predicts a woman is given byHis_iron, the probability that the network predicts a male is given byHis_Goodand the probability that the network predicts agelostrepresentHis_dob(lost), Then:

$T_{F E}=\sum_{i=1}^{11} T_i$ (11)

$T_{M A}=\sum_{i=12}^{22} T_i$ (12)

$T_{a g e}(l)=T_l+T_{l+11}$ (13)

in text,SoFunction range from 1 toSo, and gender uses 0 (feminine) andSo+1 (male). Then add the face function value and gender value to get (So+1)*2 dimensional output space. For women (gender value 0) own values ​​are importantIis the probability of the outcomeMecca liverlayer. For men (gender valueSo+1), the probability of a significant eigenvalueII (SoOutput +1+i)-totalMecca liverlayer. Therefore, the overall probability of having a significant facial feature value isIis the probability of both the output and (So+1+I)-th outputMecca liverlayer. To obtain classification probability values ​​for all significant facial features, this procedure can be applied to all significant feature values ​​(from 1 toSoFirst, initialize the array of probability valuesPlengthSo.For each significant eigenvalue, calculateP[I]=Mecca liver_Exit[I]+Mecca liver_Exit[So+1+I].A string at the endPwill contain probability values ​​for all classes of relevant facial features.

In network optimization, the gender loss function and the facial expression loss function are cross-entropy loss functions,big1 loss function. In the salient facial feature classification task, this article treats salient facial feature classification as a probability distribution problem, uses the Gaussian distribution to represent the salient facial feature probability distribution, and introduces prior knowledge which is the fixed Gaussian distribution. passedKuala Lumpurdiscord. By comparing the probability distribution of salient facial features predicted by the model with a fixed Gaussian distribution, the difference between the two distributions is minimized and the prediction accuracy is improved. At the same time,Kuala LumpurThe divergence loss function has good interpretability, making the model's performance in the forecasting process easier to understand and evaluate. Suppose the true distribution is expressed asHis(A), the predictive distribution is expressed asw(A), the degree of fit between the predicted distribution and the actual distribution is expressed asC_{Kuala Lumpur}, the following formula gives the formula for calculating the KL divergence:

$C_{K L}(t \| w)=\sum_{i=1}^M t\left(a_i\right) \log \left(\frac{t\left(a_i\right)}{w\left (a_i\右)}\右)$ (14)

if we saybig1 The loss function is expressed asstrata₁,Kuala LumpurThe divergence loss function is expressed asstrata₂, the entropy loss function is expressed asstrata₃The loss function of the constructed multi-attribute recognition network can be calculated using the following formula:

$ L O S S=0.4 \nazwa_operatora{STRATA}_1+0.6 \mathrm{STRATA}_2+x \mathrm{STRATA}_3$ (15)

picture 1.calculation processRice_HisIRice_DIn standard weave

photo 2.Optimize your network structure

In a face mask segmentation task, the mask is usually a significantly different color than the facial skin. Facial skin usually has a strong red component and a weak blue component, while some popular mask colors such as blue have a strong blue component and a weak red component. The meaning of the normalized blue-red index (National Institute of Biodiversity) is a measure of the relative intensity of the blue and red components of the image. Therefore usingNational Institute of BiodiversityIt can help distinguish facial skin from mask areas. Indexes are defined as follows:

$N B R I=\frac{Y-S}{Y+S}$ (16)

The blue and red components are expressed asIIsmall, this is. In the mask segmentation task largerNational Institute of BiodiversityThe values ​​may indicate masked areas in the image, especially if the color of the mask is significantly different from the color of the skin, such as blue. smallerNational Institute of BiodiversityThe values ​​may represent areas of skin in the image because facial skin tends to have a higher red component and a lower blue component.

In the HSI color space, we can define the index of the normalized mask area (MRI) Hue based functions (H), saturation (small) and intensity (ISince the HSI color space can better reflect the human eye's perception of color, it can perform better in the task of mask segmentation. Suppose the saturation and intensity in HIS are expressed asFII, we define the normalized indices of the mask region (MRI) in the following way:

$N M R I=\frac{F-I}{F+I}$ (17)

Where,smallIIRepresents the saturation and intensity components of an image. In the mask segmentation task largerMRIThe values ​​may represent masked areas of the image because masks tend to be more saturated and less intense. smallerMRIThe values ​​may represent areas of skin in the image because facial skin tends to be more intense and less saturated.

Using the standardized blue-red index and the normalized index of masked areas, an effective segmentation of the face mask can be achieved to distinguish masked areas from exposed skin areas. The combination of these two metrics can provide better segmentation results under different color and brightness conditions, especially when the HSI color space can better distinguish between mask and skin features.

In this article, the attenuation algorithm is combined with the enhanced grayscale world algorithm to compensate the face mask area and increase the difference between the face mask area and other areas of the face. The suppression algorithm helps reduce the interference of high-frequency noise and irrelevant features when segmenting the face mask, making the edges and shapes of the mask more visible. The enhanced Gray World algorithm can automatically adjust the color balance and contrast of the image under different lighting conditions, making it easier to recognize the difference between masks and skin. This compensation method can effectively improve the segmentation accuracy of the face mask under different lighting conditions and mask colors.

First, the input image is pre-processed using a damping algorithm. Suppression algorithms can suppress high-frequency noise and irrelevant image elements, thus preserving more pronounced structural features. In a mask area segmentation task, this can help highlight the edges and shape features of the mask. A modified gray world algorithm is then applied to balance the color and increase the contrast of the blanked image. Improved grayscale world algorithm adjustmentsRGBComponents of the image so that the average value is close to the set gray value to achieve automatic white balance. This approach helps to reduce the effect of changing lighting conditions, making color differences between the mask and the skin more visible.

Divide the face image into masked areasR(A,B) and unmasked regionsNS(A,B) image by suppressing the blue componentR(A,B), then estimate the illumination intensity and color of the masked and unmasked areas. The following formula shows the attenuation of the blue component:

$Y^{\prime}=\muY$ (18)

Suppose the lighting color of the mask area is expressed asbig(R), the lighting color of the unmasked area is expressed asbig(Pan), the image of the mask area is expressed asR(A,B), the image of the unmasked area is expressed asPan(A,B), the constant constant is expressed asX, the exponent parameter in the Minkowski norm is expressed asHisAccording to the calculation results, the mask area is compensated as follows:

$l(R)=x\left\{\frac{\iint\left(R^{\hat{\częściowy}}(a, b)^t d a d b\right)}{\iint d a d b}\right\}^ {\frac{1}{t}}$ (19)

$l(M R)=x\lijevo\{\frac{\iint\lijevo(M R^{\partial}(a, b)^t d a d b\desno)}{\iint d a d b}\right\}^{\frac{ 1}{t}}$ (20)

Correct the image of the mask area according to the following formula to achieve the purpose of mask compensation:

$\bar{R}(a, b)=R(a, b) \times l(M R) / l(R)$ (21)

To achieve a separation of face mask and non-mask areas in an image, this article uses a geometrically weighted linkage analysis model to analyze potential face mask areas. The model takes into account both geometric and textural data of the face mask area, which improves the ability to characterize features of different regions. First, the local image variance is extracted as texture information. The texture data can reflect the change in pixel values ​​in the local area, which helps distinguish the characteristics of different areas. Geometric features are calculated using an enhanced aspect ratio that helps measure area shape features and better distinguish between masked and unmasked areas. Combining texture information and geometric features, a geometrically weighted joint analysis model was constructed. The mask candidate areas are analyzed and the final mask area is determined based on the results of the geometrically weighted link analysis model. Suppose the length of the diagonal of the bounding rectangle of the connected region object is expressed asPotassium, the area of ​​the objects of the combined region is expressed asR, the largest connected area of ​​the object is expressed asX_maximum, area and linked object are expressed asX_I, the variance and the related object are expressed aselectronic_I.The model must reach the thresholdP₁IP₂.

$H Q=\left(M i>P_1\right) \cup\left(P o

$M i=\left(K^2 / R\right)$ (23)

$P o=\frac{X_{\max }}{x_i} \cdot \varepsilon_i$ (24)

The following are the technical steps of facial mask extraction based on spectral features:

(1) Pre-processing: Noise removal and edge enhancement is performed on the input face image to improve image quality and create a more accurate basis for the next steps.

(2) Obtain the mask image: use the normalized blue-red component index or other methods such as based onhiscolor space to extract potential face mask areas from the original image and generate the corresponding binary mask images.

(3) Fuzzy cluster segmentation: the mask image is segmented using the fuzzy clustering algorithm to obtain a set of possible candidate regions.

(4) Area compensation: Combining the attenuation algorithm and the enhanced gray world algorithm to compensate the face mask area, increasing the difference between the face mask area and other areas of the face for later extraction.

(5) Collecting candidate regions: analysis of the compensated image to obtain candidate regions that may contain masking regions.

(6) Geometrically weighted linkage analysis: The geometrically weighted linkage analysis model is used to analyze candidate regions, taking into account both geometric and textural information. The local variance is used as a texture feature, and the enhanced aspect ratio is used as a geometric feature to calculate area weight. Comprehensive analysis of these features identifies the actual areas of the mask on the face and eliminates unmasked areas. Figure 3 shows the technical block diagram of the algorithm proposed in this article.

photo 3.Technical block diagram of the algorithm in this article

Looking at the results of the grayscale histogram of the saturation component and the grayscale histogram of the lightness component shown in Figures 4 and 5, we can draw the following conclusions. In the saturation component grayscale histogram, most pixels are concentrated between 135 and 140, indicating that the saturation of the face mask area is relatively low and the color relatively uniform. In the grayscale histogram of the luminance component, most pixels are concentrated between 0 and 50, and the distribution of pixels in other areas is almost zero. This shows that the distribution of the mask area in the H component is relatively narrow with little color variation. Combining these observations, we can analyze that by combining the attenuation algorithm and the enhanced gray world algorithm, the color difference between the face mask area and other areas can be improved while preserving the original texture and edge information, thus improving the accuracy and robustness of recognition Viscosity. Divided mask.

In addition, the article compares and analyzes the experimental results of the face detection method optimization strategy as shown in Table 1.Meow, IpAcc.OriginalVGGThe -11 network performed well in terms of feature extraction accuracy and gender recognition accuracy, but there is still room for improvement. compared to the originalVGG-11 network, deleteConvert8,FC, IFCLayer 2 improves the accuracy of gender recognition andMeowBut this slightly reduces the accuracy of feature extraction andpAccThis suggests that reducing network complexity helps gender recognition and segmentation efficiency to some extent. Resizing the input image in this optimization strategy greatly improves gender recognition accuracypAcc, but slightly reduces the accuracy of extracting features iMeowThis shows that resizing the input image can help improve gender recognition accuracy. adjust sizeFC3 layers greatly improved the accuracy of feature extraction and gender recognition, and increasedMeowIpAcc.indicating that the resizingFCLayer 3 can effectively improve network performance. Combine highlights with gender to improve feature extraction accuracy, gender recognition accuracy,Meow, IpAccThis proves that combining salient features with gender helps improve network performance. compared to the originalVGG-11 networks, the final model is about feature extraction accuracy, gender recognition accuracy,Meow, IpAccThis shows that the integration of the above optimization strategies can effectively improve the performance of face detection methods. Comparing the different optimization strategies and their results, it can be concluded that the face detection method optimization strategy combined with significant features and gender constraints effectively improves feature extraction accuracy, gender recognition accuracy,Meow, IpAccfor more accurate facial recognition.

Figure 4.Histogram of saturation components of the mask area

Figure 5.Histogram of the brightness components of the mask area

Table 1.Comparative analysis of the results of experimental strategies for optimizing face detection methods

Table 2.Experimental results of the multi-attribute recognition network loss function

Table 3.Comparison of experimental results of different mask surface compensation methods

Table 4.Comparison of experimental results of different face mask surface segmentation methods (unit)

Table 2 shows the experimental results of the multi-attribute recognition network loss function. From the table, it can be seen that the different loss functions have a great influence on the accuracy of feature extraction, gender recognition accuracy,Meow, IpAcc.OriginalVGGThe -11 network performed well in terms of feature extraction accuracy and gender recognition accuracy, but there is still room for improvement. compared to the originalVGG-11 Networks, Introductionbig1 The loss function strategy significantly improves the accuracy of feature extraction,Meow, IpAcc, but slightly reduces gender recognition accuracy. It showsbig1 The loss function may improve the accuracy of feature extraction but has little effect on gender recognition. The introduction of the cross-entropy loss function greatly improves the accuracy of feature extraction and sex recognition, but decreases slightlyMeowIpAccThis shows that the entropy loss function has a significant impact on feature extraction and gender recognition. currentKuala LumpurThe divergence loss function greatly improves the accuracy of feature extraction, gender recognition andpAccalso increasesMeow.It indicatesKuala LumpurThe divergence loss function can further improve the accuracy of feature extraction and gender recognition. By comparing the different loss functions and their results, we can conclude that a multi-attribute recognition network has been introducedbig1 loss function, entropy loss function iKuala LumpurBoth discrepancy loss features help improve feature extraction accuracy and gender recognition accuracy. In particular, the introduction of the divergence loss function KL significantly improves the overall performance of the model. Therefore, multi-attribute recognition networks with these loss functions are very effective in improving recognition accuracy.

When constructing a dataset for mask segmentation, it is necessary to ensure the diversity and representativeness of the dataset in order to train a model with a high generalization ability. This article constructs a public scene dataset and a specific scene dataset. When constructing the dataset, ensure that the sample size is sufficient and balanced across classes to avoid training errors caused by class imbalances. At the same time, when dividing the training set, validation set and test set, it is ensured that the distribution of samples in each subset is approximately the same in order to more accurately assess the performance of the model. If the face mask segmentation model is to be applied to a specific industry, more relevant scene images can be added to the dataset to train a more targeted model.

Figure 6 shows the loss accuracy curves for network training. It can be seen that the constructed network model converges quickly during the learning process and tends to be stable thereafter. The accuracy rate of recognizing prominent facial features and gender reached 92%, and later exceeded 96%. These metrics show that the network model is performing well. Combining this information, it can be confirmed that the network model constructed in this paper has a high accuracy rate in the face mask segmentation task based on a combination of salient characteristics and gender constraints. This shows that the model can effectively identify salient facial features and gender attributes, thereby improving the accuracy of face mask segmentation. The model converges quickly during training and is stable afterwards, indicating that the model can learn efficient feature representations during training and thus provide better predictions in the testing phase.

According to Table 3, we can see the experimental results of different mask surface compensation methods on two sets of samples. It can be seen that in two sets of samples, the combination algorithm adopted in this article achieved the best segmentation performance among whichunitare 0.914 and 0.967, respectively. For other methods, the gray world algorithm, two-sided filtering, and backprojection performed relatively well on two sets of samples. This shows that these methods are robust to different types of sample sets. The Gaussian pyramid fusion and Laplace pyramid fusion methods performed poorly in sample set 1 but improved significantly in sample set 2. This may mean that these two methods are better at adapting to certain image types in sample set 2, but their ability to generalization is poor. In sample set 1, the histogram equalization and color balance methods give relatively stable results over two sample sets, but the overall performance is not as good as the other methods. This suggests that these methods may not be good enough for the mask area segmentation task.

According to Table 4, we can see the experimental results of different face mask area segmentation methods on two sets of samples. In two sets of samples, the segmentation effect of the model proposed in this paper is the best, zunitThe values ​​are 0.8187 and 0.9525 respectively. Among other methods, fuzzy c-means clustering, edge detection, and leveling methods work relatively well on both sets of samples. This shows that these methods are robust to different types of sample sets. Threshold Segmentation, Graph Slicing Algorithm andyouwangPerforms poorly on sample set 1, but improves significantly on sample set 2. This may mean that these methods have better adaptability to some types of images in sample set 2, but weaker ability to generalize in sample set 1. Although the proposed model achieves good results on both sets of samples, there is still room for further optimization to improve its suitability in different scenarios. In conclusion, face mask area segmentation experiments on different types of sample sets can clearly show that the proposed model has better performance and greater generalizability.

Figure 6.Network loss curve and accuracy