Face detection method of main sample attention mechanism
A technology of face detection and attention, applied in the field of face detection, to achieve the effect of reducing false detection, increasing detection performance, and improving 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 primary sample attention, including the following steps:
[0064] Step S100: Collect the image data containing the face in the natural scene, constitutes a training set;
[0065] Step S200: Using the network model to extract the feature map information of the training image, then by the anchor frame filter module to generate a predicted anchor frame set using the predicted coordinate value offset, and divide the prediction anchor frame into candidate Positive samples and candidate negative samples, sorting candidate positive samples and candidate negative samples, respectively;
[0066] Step S300: Finally, the loss function is used to calculate the loss value. The loss function is divided into a category classification loss function and the position regression loss function, and gives a different degree of loss value weights depending on the candidate positive sample and the ranking of candidate negative samples. The sample attention mechanism w...
Embodiment 2
[0079] This embodiment is optimized on the basis of the first embodiment, and the class classification loss function uses a binary cross entropy loss function, and the loss value weight is set according to the sample in the set of samples, which is different importance. The loss weight of the sample assignment, the calculation formula is as follows:
[0080]
[0081] among them:
[0082] It is sorted after normalization.
[0083] The maximum value ranked in the collection;
[0084] Ranking for predicting the anchor box in the collection;
[0085] To convert the weight value of the collections;
[0086] β and γ are parameters of setting adjustment weight values, respectively;
[0087] will Normalize ,
[0088] Tags for real categories,
[0089] To predict the probability value of a certain type,
[0090] Equation (4) i For the positive sample, j For negative samples,
[0091] If the ratio between the predictive anchor frame and the real boundary box is greater than ...
Embodiment 3
[0094] This embodiment is optimized based on the basis of Example 1 or 2, and the position regression loss function uses a classified sensible regression loss function, and the classification and positioning regression is positively correlated, and the classification and position regression branches are optimized, and the prediction anchor frame is calculated. The loss value between the coordinate offset and the real coordinate offset, the calculation formula is as follows:
[0095]
[0096] among them:
[0097] For universal smooth L1 loss functions,
[0098] The category probability value after the adjustment factor is calculated.
[0099] n To the category probability The total number of samples,
[0100] For the returns loss weight value obtained after normalization according to the category probability value,
[0101] For real coordinate offset,
[0102] To predict the coordinate offset;
[0103] The category probability value of the anchor frame,
[0104] b wi...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


