A Face Detection Method Based on Master Sample Attention Mechanism
A technology of face detection and attention, which is applied in the field of face detection to achieve the effect of improving accuracy and performance
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0063] A face detection method of a main sample attention mechanism, comprising the following steps:
[0064] Step S100: collecting image data containing human faces in natural scenes to form a training set;
[0065] Step S200: Use the network model to extract the feature map information of the training image, and then use the anchor frame screening module to generate a predicted anchor frame set using the predicted coordinate value offset on each feature point of the feature map, and classify the predicted anchor frames into candidate Positive samples and candidate negative samples, respectively sorting candidate positive samples and candidate negative samples;
[0066] Step S300: Finally, use the loss function to calculate the loss value. The loss function is divided into category classification loss function and position regression loss function, and according to the ranking results of candidate positive samples and candidate negative samples, the weights of loss values w...
Embodiment 2
[0079] This embodiment is optimized on the basis of Embodiment 1. The category classification loss function uses a binary cross-entropy loss function, and the weight of the loss value is set according to the order of the samples in the set, which is different importance The loss weight corresponding to the sample distribution, the calculation formula is as follows:
[0080]
[0081] in:
[0082] is the sorting after normalization,
[0083] is the maximum value ranked in the set;
[0084] To predict the ranking of the anchor box in the set to which it belongs;
[0085] is the weight value converted into the ranking of the set;
[0086] β and γ are the parameters for setting the adjustment weight value respectively;
[0087] Will normalized into ,
[0088] is the ground truth class label,
[0089] To predict the probability value of a certain class,
[0090] in formula (4) i is a positive sample, j is a negative sample,
[0091] If the distance inter...
Embodiment 3
[0094] This embodiment is optimized on the basis of Embodiment 1 or 2. The position regression loss function adopts the classification-aware regression loss function, which forms a positive correlation between classification and position regression, optimizes the classification and position regression branches at the same time, and calculates the predicted anchor frame The loss value between the coordinate offset and the real coordinate offset, the calculation formula is as follows:
[0095]
[0096] in:
[0097] is a general smooth L1 loss function,
[0098] is the category probability value after calculation with the adjustment factor,
[0099] n is the class probability value The total number of samples,
[0100] is the regression loss weight value obtained after normalization according to the category probability value,
[0101] is the real coordinate offset,
[0102] is the predicted coordinate offset;
[0103] is the category probability value of t...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More