A method and system for identifying hair loss levels

Abstract

The present invention discloses a method and system for identifying hair loss levels, and the method comprises the following steps: S1: obtaining an image including a face, and training a multi-scale convolutional layer for face detection, and extracting the face area in the image; S2: preprocessing the obtained face area image; S3: inputting the preprocessed face area image into a pre-trained hair loss level recognition module for recognition, and outputting the hair loss level. The present invention adopts a low-cost and lossless image acquisition method. Compared with the method of going to the hospital to see a doctor, the present invention has a lower time cost and higher feasibility. It adopts a deep learning method and has a better hair loss level recognition accuracy. It adopts a machine recognition method. Among shy teenagers, the present invention is more easily accepted by teenagers than the method of going to the hospital to see a doctor.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and more specifically, to a method and system for identifying hair loss levels. Background Art

[0002] Hair loss is becoming more and more common among teenagers. Sometimes, due to psychological reasons, they are too shy to see a doctor or tend to ignore the hair loss, which makes the hair loss problem more and more serious. Through the classification of hair loss levels, you can clearly understand the extent of your hair loss. The current classification of hair loss levels is mainly determined by doctors. This method requires time to go to the hospital, and some patients find it difficult to take time to go to the hospital.

[0003] A Chinese patent with a publication date of October 17, 2012 and a publication number of CN102727181A discloses a method for measuring the degree of hair loss by using a mathematical algorithm. Based on the number of hairs, hair thickness and hair follicles of the subject, a numerical value reflecting the degree of hair loss is calculated by a mathematical algorithm. Compared with the existing method of measuring the degree of hair loss based on the experience and visual observation of the measurer, the method has the advantage of being able to more accurately measure the degree of hair loss. In addition, by accurately measuring the degree of hair loss, appropriate treatment can be carried out according to the degree of hair loss, thereby having the advantage of being able to effectively treat hair loss. However, this method still requires a series of medical measurements, which is very inconvenient for patients. Summary of the invention

[0004] The primary purpose of the present invention is to provide a method for identifying hair loss levels, which provides convenience for patients. Patients can directly use the present invention to check their own hair loss levels.

[0005] A further object of the present invention is a hair loss level identification system.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A method for identifying hair loss levels comprises the following steps:

[0008] S1: obtaining an image including a face, and training a multi-scale convolutional layer for performing face detection and extracting a face region in the image;

[0009] S2: preprocessing the obtained face area image;

[0010] S3: Input the preprocessed face area image into a pre-trained hair loss level recognition module for recognition, and output the hair loss level.

[0011] Preferably, the image including the face is obtained in step S1, and the specific method of obtaining the image is as follows:

[0012] The images including human faces are automatically collected by a collection device including a camera or a still camera or are manually uploaded by a user, and the images including human faces include frontal face images and top head images of the person.

[0013] Preferably, the training of a multi-scale convolutional layer is specifically as follows:

[0014] In the training phase, the relevant face data of the public database CASIA-WebFace face recognition data set is used for training, and its loss function is:

[0015]

[0016] In the formula, K represents the length of the grid, such as K is 13 in Y1, K is 26 in Y2, and K is 52 in Y3. M represents the M different areas in a grid. Indicates whether the jth region of the i-th grid represents an object, if yes, it is 1, otherwise it is 0. Indicates whether the jth region of the i-th grid does not represent an object, if yes, it is 1, otherwise it is 0, is the actual center point of the region, (xi, yi) is the predicted center point of the region, is the actual length and width of the region, (wi,hi) is the predicted length and width of the region, is the confidence that the target is contained in the area, Ci is the confidence of the prediction, is the true probability that the area belongs to the C-type object, P i (c) is the predicted probability, λ coord， λ noobj is the weight coefficient used to balance the contribution of different loss terms.

[0017] Preferably, extracting the face area in the image in step S1 is specifically:

[0018] Through the trained multi-scale convolutional layer recognition, the grid predicted to be a human face is extracted, and the corresponding center coordinates (x, y) and length and width (w, h) are obtained. The face area is segmented from the original image according to the coordinates and length and width.

[0019] Preferably, in step S1, a multi-scale convolutional layer is trained for face detection as follows:

[0020] Inputting the image including the face into the first MCBL*2 layer, the first residual*1 layer, the first MCBL*1 layer, the first residual*2 layer, the first MCBL*1 layer and the first residual*8 layer in sequence;

[0021] The output of the first residual*8 layer is sequentially input to the first CBL*1 layer, the second residual*8 layer, the second CBL*1 layer, the first residual*4 layer, the first CBL*5 layer, the third CBL*1 layer, and the first Conv layer to obtain the feature map Y1;

[0022] The output of the first residual*4 layer is passed through the fourth CBL*1 layer and up-sampled, and then the output of the second residual*8 layer is input into the second CBL*5 layer, the fifth CBL*1 layer, and the second Conv layer in sequence to obtain the feature map Y2;

[0023] The output of the second CBL*5 layer is upsampled by the sixth CBL*1 layer, and then input into the third CBL*5 layer, the seventh CBL*1 layer, and the third Conv layer in sequence with the output of the first residual*8 layer to obtain the feature map Y3;

[0024] Based on the feature maps Y1, Y2 and Y3, a face area in the image including the face is framed.

[0025] Among them, residual represents residual block, Conv represents convolution layer, CBL represents convolution module. A convolution module consists of convolution layer Conv, normalization layer BN and leaky linear unit activation layer Leakyrelu. MCBL represents multi-scale convolution module. A multi-scale convolution module consists of multiple convolution layers of different scales, normalization layers and leaky linear unit activation layers.

[0026] Preferably, the MCBL is a multi-scale convolution, which is composed of a plurality of convolutional layers of different scales, a normalization layer and a leaky linear unit activation layer. The operation of the multi-scale convolution is as follows:

[0027]

[0028] f i =concat(f i 1*1 , f i 3*3 , f i 5*5 )

[0029] In the formula, They are 1*1 convolution kernel, 3*3 convolution kernel, and 5*5 convolution kernel, respectively, connected to the j-th output feature map of the l-1 layer And the i-th output feature map f of the multi-scale convolutional layer i 1*1 、f i 3*3 、f i 5*5 , is the bias, n is the number of output feature maps of the l-1 layer, Leakyrelu() is the leaky rectified linear unit activation function; finally, concat() is the concatenation operation, which concatenates the three feature maps in the channel dimension. The specific operation is:

[0030] If the sizes of the original three feature maps are w*h*c 1 ,w*h*c 2 ,w*h*c 3 , then the final spliced ​​output is w*h*(c 1 +c 2 +c 3 ).

[0031] Compared with the traditional YOLO V3 model, a multi-scale convolution module is used to replace the original convolution module, which can capture information of different scales on the original image, thereby obtaining more accurate face image recognition. The input image is resized to 416×416×3 and then input into the network. After a series of convolution layers, residual structures, and tensor splicing in the model, the convolution layers and tensor splicing are similar to the operations of the multi-scale convolution module mentioned above. The residual connection is to add the original feature map to the output feature map, such as:

[0032] f(x) = concat(conv(x), x)

[0033] Among them, conv(x) represents the feature map after convolution, and concat() represents tensor concatenation.

[0034] Preferably, in step S2, the obtained face region image is preprocessed, specifically:

[0035] After the face area is obtained in step S1, it is segmented from the original image, and the image of the face part is adjusted to the size that meets the input of the convolutional neural network, specifically 224×224×3. The image is then normalized, that is, the original image is divided by 255, thereby speeding up the training speed and improving accuracy and reliability.

[0036] Preferably, the hair loss level recognition module in step S3 adopts a VGG16 network model for recognition, and the VGG16 network model includes a convolutional layer, a fully connected layer and a maximum pooling layer, wherein the last fully connected layer is a classification layer, and the number of neurons in the last layer is defined according to actual task requirements, specifically:

[0037] The 224×224×3 image preprocessed in step S2 is input into the network. After two 3×3 convolutional layers and the re1u layer, it becomes a 224×224×64 matrix. Then it passes through a 2×2 maximum pooling layer with a stride of 2 to reduce the height and width to half of the original, becoming a 112×112×64 matrix. After passing through the remaining convolutional layers and pooling layers, the feature map is sent to the subsequent full connection and softmax layers to obtain the output vector. The loss function uses the cross entropy loss function:

[0038]

[0039] Among them, K is the number of categories, y is the label, if the category of the sample is i, then y i =1, otherwise y i is equal to 0, p i is the probability of category i output by the classification model.

[0040] Preferably, when the hair loss level recognition module is trained, relevant data sets are obtained as training data, and the data is enhanced online during training, such as rotation, translation, adding noise, etc., and then transfer learning is used to use the vgg16 weights trained with the imagenet data set as the initial weights of the network, and then the last classification layer is replaced, the weights of the convolutional layer are fixed, the classification layer is iterated several times, and finally all the layers are fine-tuned, and finally a convolutional neural network model that can distinguish different hair loss levels can be obtained.

[0041] The present invention also discloses a hair loss level identification system, comprising:

[0042] An image acquisition module, the image acquisition module is used to acquire an image including a face, and train a multi-scale convolutional layer for performing face detection and extracting a face area in the image;

[0043] A preprocessing module, wherein the preprocessing module is used to preprocess the obtained face area image;

[0044] A hair loss level recognition module is used to recognize the pre-processed face area image and output the hair loss level.

[0045] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0046] The present invention adopts a low-cost and lossless image acquisition method. Compared with going to the hospital to see a doctor, the present invention has lower time cost and higher feasibility. It adopts a deep learning method and has better hair loss level recognition accuracy. It adopts a machine recognition method. Among shy teenagers, the present invention is more easily accepted by teenagers than going to the hospital to see a doctor. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0048] Figure 2 It is a schematic diagram of the system module of the present invention.

[0049] Figure 3 Schematic diagram of the multi-scale convolutional layer structure of an embodiment.

[0050] Figure 4 Schematic diagram of the structure of the hair loss level recognition model in the embodiment.

[0051] Figure 5 A schematic diagram of hair loss grade results provided in an embodiment.

[0052] Figure 6 The hair loss grade standard provided for the embodiment. DETAILED DESCRIPTION

[0053] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0054] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0055] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0056] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0057] Example 1

[0058] This embodiment provides a method for identifying the level of hair loss. Figure 1 As shown, the following steps are included:

[0059] S1: obtaining an image including a face, and training a multi-scale convolutional layer for performing face detection and extracting a face region in the image;

[0060] S2: preprocessing the obtained face area image;

[0061] S3: Input the preprocessed face area image into a pre-trained hair loss level recognition module for recognition, and output the hair loss level.

[0062] Preferably, the image including the face is obtained in step S1, and the specific method of obtaining the image is as follows:

[0063] The images including human faces are automatically collected by a collection device including a camera or a still camera or are manually uploaded by a user, and the images including human faces include frontal face images and top head images of the person.

[0064] Preferably, the training of a multi-scale convolutional layer is specifically as follows:

[0065] In the training phase, the relevant face data of the public database CASIA-WebFace face recognition data set is used for training, and its loss function is:

[0066]

[0067] In the formula, K represents the length of the grid, such as K is 13 in Y1, K is 26 in Y2, and K is 52 in Y3. M represents the M different areas in a grid. Indicates whether the jth region of the i-th grid represents an object, if yes, it is 1, otherwise it is 0. Indicates whether the jth region of the i-th grid does not represent an object, if yes, it is 1, otherwise it is 0, is the actual center point of the region, (xi, yi) is the predicted center point of the region, is the actual length and width of the region, (wi,hi) is the predicted length and width of the region, is the confidence that the target is contained in the area, Ci is the confidence of the prediction, is the true probability that the area belongs to the C-type object, P i (c) is the predicted probability, λ coord ,λ noobj is the weight coefficient used to balance the contribution of different loss terms.

[0068] Preferably, extracting the face area in the image in step S1 is specifically:

[0069] Through the trained multi-scale convolutional layer recognition, the grid predicted to be a human face is extracted, and the corresponding center coordinates (x, y) and length and width (w, h) are obtained. The face area is segmented from the original image according to the coordinates and length and width.

[0070] Preferably, in step S1, a multi-scale convolutional layer is trained for face detection, such as Figure 3 , specifically:

[0071] Inputting the image including the face into the first MCBL*2 layer, the first residual*1 layer, the first MCBL*1 layer, the first residual*2 layer, the first MCBL*1 layer and the first residual*8 layer in sequence;

[0072] The output of the first residual*8 layer is sequentially input to the first CBL*1 layer, the second residual*8 layer, the second CBL*1 layer, the first residual*4 layer, the first CBL*5 layer, the third CBL*1 layer, and the first Conv layer to obtain the feature map Y1;

[0073] The output of the first residual*4 layer is passed through the fourth CBL*1 layer and up-sampled, and then the output of the second residual*8 layer is input into the second CBL*5 layer, the fifth CBL*1 layer, and the second Conv layer in sequence to obtain the feature map Y2;

[0074] The output of the second CBL*5 layer is upsampled by the sixth CBL*1 layer, and then input into the third CBL*5 layer, the seventh CBL*1 layer, and the third Conv layer in sequence with the output of the first residual*8 layer to obtain the feature map Y3;

[0075] Based on the feature maps Y1, Y2 and Y3, a face area in the image including the face is framed.

[0076] Among them, residual represents residual block, Conv represents convolution layer, CBL represents convolution module. A convolution module consists of convolution layer Conv, normalization layer BN and leaky linear unit activation layer Leakyrelu. MCBL represents multi-scale convolution module. A multi-scale convolution module consists of multiple convolution layers of different scales, normalization layers and leaky linear unit activation layers.

[0077] Preferably, the MCBL is a multi-scale convolution, which is composed of a plurality of convolutional layers of different scales, a normalization layer and a leaky linear unit activation layer. The operation of the multi-scale convolution is as follows:

[0078]

[0079] f i =concat(f i 1*1 , f i 3*3 , f i 5*5 )

[0080] In the formula, They are 1*1 convolution kernel, 3*3 convolution kernel, and 5*5 convolution kernel, respectively, connected to the j-th output feature map of the l-1 layer And the i-th output feature map f of the multi-scale convolutional layer i 1*1 、f i 3*3 、f i 5*5 , is the bias, n is the number of output feature maps of the l-1 layer, Leakyrelu() is the leaky rectified linear unit activation function; finally, concat() is the concatenation operation, which concatenates the three feature maps in the channel dimension. The specific operation is:

[0081] If the sizes of the original three feature maps are w*h*c 1 ,w*h*c 2 ,w*h*c 3 , then the final spliced ​​output is w*h*(c 1 +c 2 +c 3 ).

[0082] Compared with the traditional YOL0 V3 model, a multi-scale convolution module is used to replace the original convolution module, which can capture information of different scales on the original image, thereby obtaining more accurate face image recognition. The input image is resized to 416×416×3 and then input into the network. After a series of convolution layers, residual structures, and tensor splicing in the model, the convolution layers and tensor splicing are similar to the operations of the multi-scale convolution module mentioned above. The residual connection is to add the original feature map to the output feature map, such as:

[0083] f(x) = concat(conv(x), x)

[0084] Among them, conv(x) represents the feature map after convolution, and concat() represents tensor concatenation.

[0085] Preferably, in step S2, the obtained face region image is preprocessed, specifically:

[0086] After the face area is obtained in step S1, it is segmented from the original image, and the image of the face part is adjusted to the size that meets the input of the convolutional neural network, specifically 224×224×3. The image is then normalized, that is, the original image is divided by 255, thereby speeding up the training speed and improving accuracy and reliability.

[0087] Preferably, the hair loss level recognition module in step S3 adopts a VGG16 network model for recognition, and the VGG16 network model includes a convolutional layer, a fully connected layer and a maximum pooling layer, wherein the last fully connected layer is a classification layer, and the number of neurons in the last layer is defined according to actual task requirements, specifically:

[0088] like Figure 4, the 224×224×3 image preprocessed in step S2 is input into the network. After two 3×3 convolutional layers and the relu layer, it becomes a 224×224×64 matrix. Then, it passes through a 2×2 maximum pooling layer with a stride of 2 to reduce the height and width to half of the original, becoming a 112×112×64 matrix. After the remaining convolutional layers and pooling layers, the feature map is sent to the subsequent full connection and softmax layers to obtain the output vector. The loss function uses the cross entropy loss function:

[0089]

[0090] Among them, K is the number of categories, y is the label, if the category of the sample is i, then y i =1, otherwise y i is equal to 0, p i is the probability of category i output by the classification model.

[0091] Preferably, when the hair loss level recognition module is trained, relevant data sets are obtained as training data, and the data is enhanced online during training, such as rotation, translation, adding noise, etc., and then transfer learning is used to use the vgg16 weights trained with the imagenet data set as the initial weights of the network, and then the last classification layer is replaced, the weights of the convolutional layer are fixed, the classification layer is iterated several times, and finally all the layers are fine-tuned, and finally a convolutional neural network model that can distinguish different hair loss levels can be obtained.

[0092] In the specific implementation process, Figure 5 , 6 As shown, the current accuracy of the classification of bald attributes based on the CelebA dataset is 0.9789, the recall rate of the bald category is 0.9598, and the recall rate of the non-bald category is 0.9793, which shows the feasibility of this technology.

[0093] Example 2

[0094] This embodiment provides a hair loss level identification system. Figure 2 , used to implement the hair loss level identification method described in Example 1, comprising:

[0095] An image acquisition module, the image acquisition module is used to acquire an image including a face, and train a multi-scale convolutional layer for performing face detection and extracting a face area in the image;

[0096] A preprocessing module, wherein the preprocessing module is used to preprocess the obtained face area image;

[0097] A hair loss level recognition module is used to recognize the pre-processed face area image and output the hair loss level.

[0098] The same or similar reference numerals correspond to the same or similar components;

[0099] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;

[0100] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for identifying hair loss levels, characterized in that: The following steps are involved: S1: obtaining an image including a face, and training a multi-scale convolutional layer for performing face detection and extracting a face region in the image; S2: preprocessing the obtained face area image; S3: input the preprocessed face area image into a pre-trained hair loss level recognition module for recognition, and output the hair loss level; In step S1, a multi-scale convolutional layer is trained for face detection as follows: Inputting the image including the face into the first MCBL*2 layer, the first residual*1 layer, the first MCBL*1 layer, the first residual*2 layer, the first MCBL*1 layer and the first residual*8 layer in sequence; The output of the first residual*8 layer is sequentially input to the first CBL*1 layer, the second residual*8 layer, the second CBL*1 layer, the first residual*4 layer, the first CBL*5 layer, the third CBL*1 layer, and the first Conv layer to obtain the feature map Y1; The output of the first residual*4 layer is passed through the fourth CBL*1 layer and up-sampled, and then the output of the second residual*8 layer is input into the second CBL*5 layer, the fifth CBL*1 layer, and the second Conv layer in sequence to obtain the feature map Y2; The output of the second CBL*5 layer is upsampled by the sixth CBL*1 layer, and then input into the third CBL*5 layer, the seventh CBL*1 layer, and the third Conv layer in sequence with the output of the first residual*8 layer to obtain the feature map Y3; Frame a face region in the image including the face according to the feature maps Y1, Y2 and Y3; The training of a multi-scale convolutional layer is specifically as follows: In the training phase, the relevant face data of the public database CASIA-WebFace face recognition data set is used for training, and its loss function is: In the formula, K represents the length of the grid, M represents the M different areas in a grid, Indicates whether the jth region of the i-th grid represents an object, if yes, it is 1, otherwise it is 0. Indicates whether the jth region of the i-th grid does not represent an object, if yes, it is 1, otherwise it is 0, is the actual center point of the region, (x i ,y i ) is the predicted center point of the region, is the actual length and width of the region, (w i ,h i ) are the predicted length and width of the region, is the confidence that the target is contained in the region, C i is the confidence of the prediction, is the true probability that the area belongs to the C-type object, P i (c) is the predicted probability, λ coord ,λ noobj is the weight coefficient, which is used to balance the contribution of different loss terms; In step S3, the hair loss level recognition module adopts the VGG16 network model for recognition. The VGG16 network model includes a convolutional layer, a fully connected layer and a maximum pooling layer, wherein the last fully connected layer is a classification layer. The number of neurons in the last layer is defined according to the actual task requirements, specifically: The 224×224×3 image preprocessed in step S2 is input into the network. After two 3×3 convolutional layers and the relu layer, it becomes a 224×224×64 matrix. Then it passes through a 2×2 maximum pooling layer with a stride of 2 to reduce the height and width to half of the original, becoming a 112×112×64 matrix. After passing through the remaining convolutional layers and pooling layers, the feature map is sent to the subsequent full connection and softmax layers to obtain the output vector. The loss function uses the cross entropy loss function: Where T is the number of categories and y is the label. If the category of the sample is t, then y t =1, otherwise y t is equal to 0, p t is the probability of category t output by the classification model; When the hair loss level recognition module is trained, relevant data sets are obtained as training data, and the data is enhanced online during training. Then, transfer learning is used to use the vgg16 weights trained with the imagenet data set as the initial weights of the network. Then, the last classification layer is replaced, the weights of the convolutional layer are fixed, the classification layer is trained iteratively several times, and finally all the layers are fine-tuned. Finally, a convolutional neural network model that can distinguish different hair loss levels can be obtained.

2. The method for identifying the hair loss level according to claim 1, characterized in that: The specific method of obtaining the image including the face in step S1 is as follows: The images including human faces are automatically collected by a collection device including a camera or a still camera or are manually uploaded by a user, and the images including human faces include frontal face images and top head images of the person.

3. The method for identifying the hair loss level according to claim 2, characterized in that: In step S1, the face area in the image is extracted, specifically: Through the trained multi-scale convolutional layer recognition, the grid predicted to be a human face is extracted, and the corresponding center coordinates (x, y) and length and width (w, h) are obtained. The face area is segmented from the original image according to the coordinates and length and width.

4. The method for identifying the hair loss level according to claim 3, characterized in that: The MCBL is a multi-scale convolution, which consists of a plurality of convolutional layers of different scales, a normalization layer and a leaky linear unit activation layer. The operation of the multi-scale convolution is as follows: In the formula, They are 1*1 convolution kernel, 3*3 convolution kernel, and 5*5 convolution kernel, respectively, connected to the mth output feature map of layer l-1 And the nth output feature map of the multi-scale convolutional layer is the bias, N is the number of output feature maps of the l-1 layer, Leakyrelu() is the leaky rectified linear unit activation function; finally, concat() is the concatenation operation, which concatenates the three feature maps in the channel dimension. The specific operation is: If the sizes of the original three feature maps are w*h*p1, w*h*p2, and w*h*p3 respectively, the final concatenated output is w*h*(p1+p2+p3).

5. The method for identifying the hair loss level according to claim 4, characterized in that: In step S2, the obtained face region image is preprocessed, specifically: After the face area is obtained in step S1, it is segmented from the original image, and the image of the face part is adjusted to a size that meets the input size of the convolutional neural network, specifically 224×224×3, and then the image is normalized, that is, the original image is divided by 255.

6. A hair loss level identification system, characterized in that: The system applies the hair loss level identification method according to any one of claims 1 to 5, and the system comprises: An image acquisition module, the image acquisition module is used to acquire an image including a face, and train a multi-scale convolutional layer for performing face detection and extracting a face area in the image; A preprocessing module, wherein the preprocessing module is used to preprocess the obtained face area image; A hair loss level recognition module is used to recognize the pre-processed face area image and output the hair loss level.

Images