Face acne detection method based on sampling effectiveness of enhanced ranking accuracy

Abstract

The application discloses a face acne detection method based on sampling effectiveness of enhanced sorting precision, which is applied to the technical field of computer vision and image processing, and aims at the problem of low detection precision of existing face acne detection technology. The artificial neural network of the application uses a multilevel convolutional neural network to perform feature extraction on an input image, and obtains a feature map F of the input; features are extracted from the feature map F again according to a foreground candidate frame obtained through preliminary regression, instance segmentation, classification and regression are performed, and an instance segmentation result M, a classification score P and a frame regression result B are obtained; features are extracted from the detection frame after final regression again, a sorting score T of the frame regression result B is learned, sampling is performed by an instance sampler and a background sampler to obtain training samples, and then the sorting network is used to train the sorting score, the classification score P is fused, and a final score S is obtained; and the face acne detection can be effectively performed by using the method.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision and image processing, and particularly relates to a facial acne detection technology. BACKGROUND

[0002] Target detection is a key task in the field of computer vision, and has wide applications in various fields such as autonomous driving, video surveillance, medical image analysis, etc. Its main purpose is to identify and locate specific targets from images or videos, and determine their categories. With decades of continuous development, the problem of target detection has made significant progress in various scenarios. In recent years, the rise of deep learning has driven the innovation of the field of target detection, introducing neural network-based methods such as Faster R-CNN, YOLO, CenterNet, etc. These methods use anchor boxes or anchor-free strategies, which can usually provide more accurate and reliable detection results compared to traditional methods.

[0003] Ranking loss method has been widely applied in object detection field in recent years. Chen Kean et al. adopted AP loss function instead of classification loss to solve the imbalance problem in the first stage of object detection model; Qi Qian et al. also adopted ranking score instead of classification score, and improved the effectiveness of ranking score by expanding the distance between positive and negative sample score distribution on the basis of simulating ranking score. Tan Zhiyu et al. divided the whole picture positive sample into multiple intervals according to the IoU value for sampling, which improved the sampling effectiveness of ranking loss. Reference: Chen K, Li J, Lin W, See J, Wang J, Duan L, Chen Z, He C, Zou J. Towards accurate one-stage object detection with ap-loss. In: Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2019, 5119-5127; Qian Q, Chen L, Li H, Jin R. Dr loss: Improving object detection by distributional ranking. In: Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2020, 12164-12172; Tan Z, Nie X, Qian Q, Li N, Li H. Learning to rank proposals for object detection. In: Proceedings of the IEEE / CVF International Conference on Computer Vision. 2019, 8273-8281;

[0004] Medical image object detection is usually more challenging than general object detection because medical objects are often difficult to distinguish from the background. Most of the medical image object detection methods are proposed for specific medical image data features, without considering the common characteristics of medical images. At the same time, due to the lack of high-quality public data sets, there are few studies on the detection of facial acne. Ethan H. Nguyen et al. used a circular detection box instead of a rectangular detection box for glomerulus and cell detection, which is more consistent with the shape of the relevant object than the rectangle; Thomas Wollmann et al. improved the network structure, anchor regression method and loss function according to the data characteristics of cells and particles in the microscope image; Yi Jing Ru et al. used an object-guided coarse-to-fine segmentation branch to exclude the interference of adjacent objects on the segmentation effect in the cell image instance segmentation task; Kuladech Rashataprucksa et al. applied deep learning object detection method to acne image and obtained better results than machine learning method; Zhang JianWei replaced the IoU branch with the NWD branch to improve the recall rate of the first stage of the two-stage model, and contributed the 10-class facial acne data set scuAcne.Nguyen E H, Yang H, Deng R, Lu Y, Zhu Z, Roland J T, Lu L, Landman B A, Fogo A B, Huo Y. Circle representation for medical object detection. IEEE transactions on medical imaging, 2021, 41(3): 746-754; Wollmann T, Rohr K. Deep consensus network: Aggregating predictions to improve object detection in microscopy images. Medical Image Analysis, 2021, 70: 102019; Yi J, Wu P, Tang H, Liu B, Huang Q, Qu H, Han L, Fan W, Hoeppner D J, Metaxas D N. Object-guided instance segmentation with auxiliary feature refinement for biological images. IEEE Transactions on Medical Imaging, 2021, 40(9): 2403-2414; Rashataprucksa K, Chuangchaichatchavarn C, Triukose S, Nitinawarat S, Pongprutthipan M, Piromsopa K. Acne detection with deep neural networks. In: 2020 2nd International Conference on Image Processing and Machine Vision. 2020, 53-56; Zhang J, Zhang L, Wang J, et al. SA-RPN: A Spacial Aware Region Proposal Network for Acne Detection [J]. IEEE Journal of Biomedical and Health Informatics, 2023.However, there are few studies on facial acne detection at present, and most of the studies on medical image target detection are based on the characteristics of specific data sets to improve traditional target detection models, and the method is difficult to generalize to other medical image data sets. SUMMARY

[0005] To solve the above technical problems, the present application provides a facial acne detection method based on enhanced sampling effectiveness of ranking accuracy, which can provide accurate position and category of the target of interest in the scene for subsequent computer applications in facial acne detection or medical image detection scenarios.

[0006] The technical scheme adopted by the present application is: a facial acne detection method based on enhanced sampling effectiveness of ranking accuracy, comprising the following steps:

[0007] S1, an artificial neural network is constructed, and the processing process of the input image by the artificial neural network is:

[0008] S11, a multi-level convolutional neural network is used to extract features from the input image to obtain a feature map F of the input;

[0009] S12, a sliding window is initialized as an initial anchor frame according to the feature map dimension, and a preliminary regression and foreground and background classification are performed on the feature map F based on the initial anchor frame, and the foreground candidate frame is screened;

[0010] S13, features are extracted from the feature map F according to the foreground candidate frame obtained in step S12, instance segmentation, classification and regression are performed, and instance segmentation result M, classification score P and bounding box regression result B are obtained;

[0011] S14, features are extracted from the feature map F using the bounding box regression result B, the ranking score T of the bounding box regression result B is learned, and the training samples are obtained by sampling with instance sampler and background sampler, then the ranking network is used to train the ranking score, and the final score S is obtained by fusing the classification score in step S3;

[0012] S2, the artificial neural network constructed in step S1 is trained using the image and annotation information provided by the data set until convergence;

[0013] S3, the input image to be detected is input into the converged artificial neural network to obtain the final detection result.

[0014] The beneficial effects of this invention are as follows: This method first uses a multi-level convolutional neural network to extract features from the input image, obtaining the input feature map F; based on the foreground candidate boxes obtained from the preliminary regression, features are extracted again from the feature map F, and instance segmentation, classification, and regression are performed to obtain the instance segmentation result M, the classification score P, and the bounding box regression result B; features are extracted again using the detection boxes after the final regression, and the ranking score T of the bounding box regression result B is learned. Samples are obtained by sampling with instance samplers and background samplers, and then the ranking network is used to train the ranking score, which is fused with the classification score P to obtain the final score S; the method of this invention can effectively detect facial acne, providing the precise location and category of the target of interest in the scene for medical image target detection applications. Attached Figure Description

[0015] Figure 1 This is a flowchart of the target detection method of the present invention;

[0016] Figure 2 This embodiment uses facial acne images and corresponding tag information.

[0017] Figure 3 This is a candidate bounding box diagram of the facial acne image obtained after step 3 in this embodiment;

[0018] Figure 4 This is a regression diagram of the facial acne image obtained after step 4 in this embodiment;

[0019] Figure 5 This is the final target detection result in this embodiment;

[0020] Figure 6 This is the final instance segmentation result in this embodiment. Detailed Implementation

[0021] To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with the accompanying drawings, further illustrates the invention.

[0022] like Figure 1 As shown, the method of the present invention includes the following implementation process:

[0023] 1. Use a multi-level convolutional neural network to extract features from the input image to obtain the input feature map F. Specifically, use a basic neural network to extract features from the input image I, and then use a feature pyramid for feature enhancement to obtain a multi-level feature map F.

[0024] F1,F2,…F i …F n =F FPN (B(I)) (1)

[0025] wherein B represents a basic convolutional neural network, which can be implemented by any commonly used convolutional neural network, such as ResNet, VGG, DenseNet, etc., to extract features of an input image, F FPN represents a feature pyramid for feature enhancement, F1~F n represents an output multi-level feature map, and the multi-level feature map F i The higher the level is, the larger the receptive field corresponding to the feature point is, and the smaller the spatial dimension of the feature map is. For example, an image with a shape of (3, 1024, 1024) is input, and after feature extraction by the basic convolutional neural network, four-level feature maps with dimensions of (256, 256, 256), (512, 128, 128), (1024, 64, 64), and (2048, 32, 32) are obtained, wherein the first dimension is the number of feature channels, and the second and third dimensions are the width and height of the feature map, respectively. Then, the multi-layer feature fusion is performed by the feature pyramid F FPN to obtain the feature maps {F1, F2, F3, F4} with dimensions of (256, 256, 256), (256, 128, 128), (256, 64, 64), and (256, 32, 32), respectively.

[0026] 2. The sliding window is initialized as an initial anchor box according to the dimension of the feature map, the initial anchor box is preliminarily regressed and classified into foreground and background, and a certain number of detection boxes are selected as foreground. Specifically, the number of sliding windows is determined according to the dimension of the multi-level feature map, and the multi-level sliding window size is designed as an initial anchor box A. The initial anchor box is preliminarily regressed and classified into foreground and background, and a certain number of regressed detection boxes are selected as foreground RoI.

[0027] P i f =F fb (ForeGroundSelector(F i ,A i ,G))(i=1,2,…n) (2)

[0028]

[0029] RoI=RoISelector(P f ,B f ) (4)

[0030] wherein F i is the i-th level feature map, A iFor the i-th level anchor box, the number of anchor boxes corresponds to the feature map dimension, and there are three kinds of anchor boxes with width-height ratios of 0.5, 1.0 and 2.0 at each feature point in the spatial dimension. According to the example above, the data dimensions of the four-level anchor boxes are (256x256x3, 4), (128x128x3, 4), (64x64x3, 4) and (1024x1024x3, 4) respectively. The last dimension of the anchor box data is the anchor box position data (center_x, center_y, w, h), that is, the center point left and the width and height values. The anchor box size can be manually designed, but is generally approximately the same as the corresponding feature map receptive field. In the example, the width and height values of the anchor box with a width-height ratio of 1.0 corresponding to the four-level feature map are (32, 32), (64, 64), (128, 128) and (256, 256) respectively. Meanwhile, in step 2, the label categories are not distinguished, and all are regarded as foreground.

[0031] ForeGroundSelector is a foreground-background sampler, and anchor boxes with an IoU value greater than or equal to 0.5 with any label G are regarded as positive samples, and the rest are regarded as negative samples. ForeGroundSelector has a fixed total number of samples sample_c for each picture, and the sampled samples are divided into foreground samples pos_c and background samples neg_c. The foreground samples are taken as much as possible but not more than that is, pos c <0.5*sample_c. The definition of the IoU value is the area of the intersection of the two rectangles divided by the area of the union of the two rectangles.

[0032] F fb 、F ar are a foreground-background discrimination network and a bounding box regression network respectively, P i f 、 are the i-th level foreground-background score and anchor box regression result respectively, such as are (256x256x3, 2) and (256x256x3, 4) respectively, and P i f The last dimension of the data is the foreground and background score of the candidate box, The last dimension of the data is the center position relative to the width-height variation multiple and the width-height variation multiple (Δx, Δy, Δw, Δh). The regression anchor box data is (center_x+w*Δx, center_y+h*Δy, w+w*Δw, h+h*Δh).

[0033] RoSelector is a foreground-background selector, and RoI is a foreground candidate box. When selecting the candidate box regression result, RoISelector first selects the candidate box with the highest foreground-background score in each level feature candidate box The first nms_pre maximum score candidate boxes are taken, and then a local optimal screening (NMS, Non-Maximum Suppression) operation is performed, and then the multi-level screening results are combined, and the max_per_image candidate boxes are taken as RoI according to the foreground score, and in the example, max_per_image = 1000, that is, the number of RoI NR≤max_per_image.

[0034] 3. Re-extract features according to the detection frame after preliminary regression, perform instance segmentation, classification and regression, and obtain instance segmentation result M, classification result P and regression result B. The specific method is: for the foreground candidate box RoI of step 2, using the feature alignment extractor to extract features:

[0035] FD = f(F, RoI) (5)

[0036] Wherein f is a feature alignment extractor, and corresponding features are extracted from the feature map F according to the RoI data using bilinear interpolation; FD is the feature map of each candidate box obtained after extraction. For the instance segmentation and target detection task, the detection sampler is used for sampling to obtain training samples, and then sample training is performed:

[0037] M = F m (DetectionSelector(FD, G)) (6)

[0038] P = F cls (DetectionSelector(FD, G)) (7)

[0039] B = F reg (DetectionSelector(FD, G)) (8)

[0040] Wherein DetectionSelector is a detection sampler, F m , F cls , F reg is an instance segmentation network, a classification network and a bounding box regression network, M, P and B are instance segmentation results, classification scores and bounding box regression results, and the data dimensions are (NR, 28, 28), (NR, 11) and (NR, 4) respectively, that is, the source data is 10 categories (plus the background class, the last dimension of the classification score is 11). For the target detection task, the detection sampler is used for sampling to obtain training samples, and then classification and bounding box regression training is performed.

[0041] DetectionSelector is similar to ForeGroundSelector, the total number of samples for each picture is fixed, and the samples are divided into foreground samples pos_d and background samples neg_d, the foreground samples are used as much as possible but not more than the total number of samples That is, pos_d < 0.25 * sample_d, and in order to prevent the number of positive samples from being too small, the label G is used as a positive sample for training. As shown in Figure 3 The candidate box diagram of the facial acne image after step 3 processing is shown.

[0042] 4, using the detection box after final regression to extract features again, learning the ranking score T of the detection box, through instance sampling and background sampling method to make the sample can be fully trained, using ranking loss function to train the ranking score of the detection box, and the background score P and the classification score T are fused to obtain the final score S. The specific method is: according to the final regression candidate box of step 3, using the feature alignment extractor f to extract accurate features again. Then use instance sampler and background sampler to sample to obtain training samples, and then use ranking network to train ranking score, and fuse with classification score in step 3 to obtain candidate box detection result.

[0043] FR = f (F, B) (9)

[0044] T = F rank ({GtAwareSelector (FR, G), BackGroundSelector (FR, G)}) (10)

[0045] Where GtAwareSelector, BackGroundSelector are instance sampler and background sampler respectively, F rank is the ranking network, and T is the ranking score output by the ranking network.

[0046] GtAwareSelector samples each instance G j respectively, for each instance matched detection box B j , according to the IoU of the detection box, the positive sample detection box with IoU ≥ 0.5 is divided into 10 detection box sets with every 0.05 Each set is traversed once, and the current set detection box score is used as a positive sample, and h detection box scores with the highest score are taken from the detection box with smaller IoU as negative samples to construct sample pairs. At the same time, through the way of instance sampling, different instance samples can be fully trained.

[0047] The BackGroundSelector takes the v highest-scored bounding boxes as negative samples for all bounding boxes with IoU < 0.5 in the bounding box B, and takes the average score of each set of 10 bounding boxes as positive samples for all bounding boxes with IoU ≥ 0.5, wherein each set of 10 bounding boxes is divided according to the IoU value at every 0.05, so that the background samples can also be effectively trained through background sampling.

[0048] The dimension of the ranking score is (NR, 1), and the final score S is obtained by fusing the ranking score T and the classification score P, as shown in the following formula:

[0049]

[0050] As shown in Figure 4 , it is a face acne image bounding box diagram after step 4 processing.

[0051] 5. Using the image and label information provided by the data set to train the artificial neural network until convergence, and then inputting the image to be detected to obtain the final detection result. The specific method is as follows: using the image and label information in the data set to train the artificial neural network until the network converges, and the specific training method is as follows:

[0052] L = λ c L c + λ m L m + λ d L d (12)

[0053] In formula (12), L represents the training loss of the neural network; L c , L m , and L d represent the foreground and background extraction loss, the instance segmentation loss, and the target detection loss, respectively. c , λ m , and λ d represent the coefficients of the foreground and background extraction loss, the instance segmentation loss, and the target detection loss, respectively. c , λ m , and λ d are all taken as 1.0 in this example.

[0054] The instance segmentation loss L m is calculated by the following formula:

[0055] L m = -(M'logM + (1-M')log(1-M)) (13)

[0056] M represents the instance segmentation map obtained from equation (6); and M' represents the supervision signal provided by the dataset.

[0057] wherein the foreground-background extraction loss L c can be calculated by the following equation:

[0058] L c = λ fb L fb + λ ar L ar (14)

[0059] In equation (14), L fb represents the foreground-background discrimination loss, and L ar represents the bounding box regression loss. λ fb , λ ar respectively represent the coefficients of the foreground-background discrimination loss and the bounding box regression loss, and λ fb , λ ar are both taken as 1.0 in this example.

[0060] The object detection loss L d can be calculated by the following equation:

[0061] L d = λ cls L cls + λ reg L reg + λ rank L rank (15)

[0062] In equation (15), L cls represents the classification loss, L reg represents the bounding box regression loss, and L rank represents the ranking loss. λ cls , λ reg and λ rank respectively represent the coefficients of the classification loss, the bounding box regression loss and the ranking loss, and λ cls and λ reg are taken as 1.0 and λ rank is taken as 0.1 in this example.

[0063] The loss functions of the classification losses L c and L cls are similar to the instance segmentation L m and cross-entropy loss is adopted. The regression losses L ar and L reg adopt L1Loss, and L reg is for example as shown in the following equation:

[0064] L reg = |B-B'| (16)

[0065] where B' is the bounding box regression label. The ranking loss L rank is calculated by the following equation:

[0066] L rank = max(0, T j -T i + a) (17)

[0067] In equation (16), T i represents the positive sample, T j represents the negative sample, and a represents the hyperparameter limiting the distance of the sample pair (T i , T j ). In this example, a = 0.1.

[0068] The loss is calculated by a specific loss function according to the model output and the label, which is provided by the facial acne dataset scuAcne labeled by the dermatologists of West China Medical College. The source data is the polygon annotation of the target, and the semantic segmentation label is directly converted from the polygon annotation, and the bounding box label is the circumscribed rectangle of the target boundary. An example of loss calculation of the neural network is as follows: the loss value is calculated by using equations (12)-(17) according to the target detection output obtained according to equations (6), (7), (8) and (10) and the label provided by the dataset, and then the weight of the entire neural network is updated using the back propagation algorithm, so as to continuously learn from the dataset, repeat epochs times, until the network converges, and the training phase ends. After the training phase is completed, the neural network weight is fixed, and the image can be input to perform the target detection task.

[0069] As shown in Figure 2 , this example uses facial acne images with a size of 1024x1024. The parameter settings in this example are as follows: epochs = 15, B = ResNet50; in the test phase, we only use the bounding box with a score greater than score_threshold = 0.3 as the result. After the calculation of steps 1-4, the final detection result is obtained by step 5, as shown in Figure 5 、 6 . Figure 2 In the above example, atrophic_scar represents atrophic scar, and papule represents papule.

[0070] The above simple example mainly uses the whole image as an example to illustrate and demonstrate, and the actual calculation is realized by corresponding local convolution and weighted summation operations on all pixels of the whole image. The actual numerical value and result are directly taken from the experimental results of program running. Through this simple example, the whole process of the facial acne detection method based on the sampling effectiveness of the enhanced bounding box ranking precision is described.

[0071] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the principles of the application and should not be construed as limiting the scope of the application to such specifically enumerated embodiments. This application can be modified and varied in various ways, and it is therefore to be understood that the application can be practiced otherwise than specifically set forth herein. Any modifications, equivalent substitutions, improvements, etc., made to the application in accordance with the spirit and principles of the application are to be embraced within the scope of the application as defined in the following claims.

Claims

1. A facial acne detection method based on improved sampling effectiveness and enhanced sorting accuracy, characterized in that, Includes the following steps: S1. Construct an artificial neural network. The processing procedure of this artificial neural network for the input image is as follows: S11. Use a multi-level convolutional neural network to extract features from the input image to obtain the input feature map F; S12. Initialize the sliding window as the initial anchor box according to the feature map dimension, perform preliminary regression and foreground / background classification on the feature map F based on the initial anchor box, and filter foreground candidate boxes. S13. Based on the foreground candidate boxes obtained in step S12, extract features from the feature map F again, perform instance segmentation, classification and regression, and obtain the instance segmentation result M, classification score P and bounding box regression result B; S14. Use the bounding box regression result B to extract features again from the feature map F, learn the ranking score T of the bounding box regression result B, sample separately for instance sampler and background sampler to obtain training samples, then use the ranking network to train the ranking score, and fuse it with the classification score in step S3 to obtain the final score S; the specific implementation process of step S14 is as follows: based on the candidate boxes after the final regression in step S13, use the feature alignment extractor Accurate features are extracted again from the multi-level feature map F; then, training samples are obtained by using the instance sampler and background sampler; then, the ranking network is used to train the ranking score, which is fused with the classification score in step S13 to obtain the candidate box detection result. The training process for ranking scores is shown in the following formula: ； ； in , These are instance samplers and background samplers, respectively. For sorting networks, The sorting score output by the sorting network; Sub-instance sampler For each instance Sampling is performed, and for each instance, a matching detection box is generated. According to the detection box Sorting is carried out on the following: The positive sample detection box, and it is set according to Value per One set is divided into 10 detection box sets. Each time, a set is traversed, and the score of the detection box in the current set is used as a positive sample. Select from a smaller set of detection boxes The detection box with the largest score is used as a negative sample to construct a sample pair; the instance-based sampling method ensures that the samples corresponding to different instances can be fully trained. Background Sampler For the detection frame All Background detection box The detection box with the largest score is used as the negative sample. The positive sample detection box, and it is set according to Value per A set is divided into 10 detection box sets. The average score of each detection box set is used as a positive sample to construct a sample pair. The final score S is obtained by fusing the ranking score T and the classification score P, as shown in the following formula: ； S2. Use the images and annotation information provided by the dataset to train the artificial neural network constructed in step S1 until it converges; S3. Input the image to be detected into the converged artificial neural network to obtain the final detection result.

2. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 1, characterized in that, Step S11 specifically involves: extracting features from the input image using a convolutional neural network and enhancing features using a feature pyramid to obtain a multi-level feature map F; the expression is: ； in Indicates the input image. This represents a basic convolutional neural network. Represents the feature pyramid used for feature enhancement. This represents the multi-level feature map output.

3. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 2, characterized in that, The specific implementation process of step S12 is as follows: determine the number of sliding windows based on the dimension F of the multi-level feature map, design the size of the multi-level sliding window as the initial anchor box A, perform preliminary regression and foreground / background classification on the initial anchor box, and select a certain number of regression detection boxes as foreground RoIs; the specific expression is as follows: ； ； ； in For the first Level feature map, For the first Level anchor frame, For target tags, Foreground and background samplers, , These are foreground / background discrimination network and bounding box regression network, respectively. , The first Level foreground and background scores and anchor frame regression results Foreground and background filters These are foreground candidate boxes.

4. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 3, characterized in that, The specific implementation process of step S13 is as follows: For the foreground candidate box RoI from step S12, features are extracted using a feature alignment extractor. ； Among them, for For feature alignment extractor, For each candidate box obtained after extraction; For instance segmentation and object detection tasks, a detection sampler is used to obtain training samples, which are then used for training; the expression is shown below: ； ； ； in To detect the sampler, , , For instance segmentation network, classification network, and bounding box regression network, , , The results are instance segmentation, classification scores, and bounding box regression results.

5. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 4, characterized in that, The loss function used in step S2 of the training process is expressed as follows: ； in This represents the training loss of the neural network; , , These represent foreground and background extraction loss, instance segmentation loss, and object detection loss, respectively. , , These represent the coefficients of the foreground / background extraction loss, instance segmentation loss, and object detection loss, respectively.

6. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 5, characterized in that, Foreground and background extraction loss The expression is: ； in Represents the foreground and background in determining loss. Represents the bounding box regression loss; , These represent the coefficients of the foreground / background discrimination loss and the bounding box regression loss, respectively.

7. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 6, characterized in that, Target detection loss The expression is: ； in Represents classification loss, Represents the bounding box regression loss. Represents the ranking loss; , and These represent the coefficients of classification loss, bounding box regression loss, and ranking loss, respectively.

8. The facial acne detection method based on enhanced sorting accuracy and sampling effectiveness according to claim 7, characterized in that, Ranking loss The expression is: ； in Represents positive samples. Representing negative samples, Representative restricted sample pairs Hyperparameters of distance.

Images