News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Face Recognition SphereFace
Face Recognition SphereFace
2017 cvpr, SphereFace: Deep Hypersphere Embedding for Face Recognition, followed by another masterpiece centerloss.
The article mainly proposes normalize weights and zero biases and angular margins. Based on these two points, the traditional softmax is improved so that the maximum intraclass distance is smaller than the smallest class interval. Distance identification criteria.
The loss loss of the original softmax is a loss of mutual entropy.
among them,
Substituting it into the first formula, the loss function is as follows,
The
Perform the normalization operation and map them to a unit sphere. Let ||W||=1, b=0, and introduce the included angle. Find the Modified Softmax Loss formula as follows.
The
Based on this, we introduce the angular margin, denoted by m, and finally produce A-softmax's loss formula
The
The recognition rates on the LFW and YTF data sets are shown in the figure below. It can be seen that the LFW has reached a recognition rate of 99.42.
Here to analyze the parameters of the author A-softmax,
In the Forward_cpu function of margin_inner_product_layer.cpp,
In other words,
Calculated that is to say, lambda = 5 (iteration> 1658) lambda = greater than 5 (iteration <1658)
The article mainly proposes normalize weights and zero biases and angular margins. Based on these two points, the traditional softmax is improved so that the maximum intraclass distance is smaller than the smallest class interval. Distance identification criteria.
The loss loss of the original softmax is a loss of mutual entropy.
among them,
Substituting it into the first formula, the loss function is as follows,
The
Perform the normalization operation and map them to a unit sphere. Let ||W||=1, b=0, and introduce the included angle. Find the Modified Softmax Loss formula as follows.
The
Based on this, we introduce the angular margin, denoted by m, and finally produce A-softmax's loss formula
The
The recognition rates on the LFW and YTF data sets are shown in the figure below. It can be seen that the LFW has reached a recognition rate of 99.42.
Here to analyze the parameters of the author A-softmax,
In the Forward_cpu function of margin_inner_product_layer.cpp,
In other words,
Calculated that is to say, lambda = 5 (iteration> 1658) lambda = greater than 5 (iteration <1658)