损失名称 | 数学公式 | 是否权重归一化 | 是否特征归一化 | 距离空间 | 分类边界 | 文章链接 |
---|---|---|---|---|---|---|
Cross Entropy Loss | L = − 1 m ∑ i = 1 m log e W y i T x i + b y i ∑ j = 1 n e W j T x i + b j L = -\frac{1}{m}\sum_{i=1}^m\log\frac{e^{W_{y_i}^Tx_i+b_{y_i}}}{\sum_{j=1}^ne^{W_j^Tx_i+b_j}} L=−m1∑i=1mlog∑j=1neWjTxi+bjeWyiTxi+byi | 否 | 否 | 欧式空间 | ( W 1 − W 2 ) x + b 1 − b 2 = 0 (W_1-W_2)x+b_1-b_2=0 (W1−W2)x+b1−b2=0 | — |
triplet loss | ∥ f ( x i a ) − f ( x i p ) ∥ 2 2 + α < ∥ f ( x i a ) − f ( x i n ) ∥ 2 2 \|f(x_i^a)-f(x_i^p)\|_2^2 + \alpha < \|f(x_i^a)-f(x_i^n)\|_2^2 ∥f(xia)−f(xip)∥22+α<∥f(xia)−f(xin)∥22 | 否 | 否 | 欧式空间 | – | facenet |
center loss | L = − 1 m ∑ i = 1 m log e W y i T x i + b y i ∑ j = 1 n e W j T x i + b j + λ 2 ∑ i = 1 m ∥ x i − c y i ∥ 2 2 L = -\frac{1}{m}\sum_{i=1}^m\log\frac{e^{W_{y_i}^Tx_i+b_{y_i}}}{\sum_{j=1}^ne^{W_j^Tx_i+b_j}}+\frac{\lambda}{2}\sum_{i=1}^m\|x_i-c_{y_i}\|_2^2 L=−m1∑i=1mlog∑j=1neWjTxi+bjeWyiTxi+byi+2λ∑i=1m∥xi−cyi∥22 | 否 | 否 | 欧式空间 | – | centerloss |
coco loss | L = ∑ i ∈ β e x p C ( f ( i ) , c l i ) ∑ m ≠ l i e x p C ( f ( i ) , c m ) L =\sum_{i\in\beta}\frac{expC(f^{(i)},c_{l_i})}{\sum_{m \neq l_i}expC(f^{(i)},c_m)} L=∑i∈β∑m=liexpC(f(i),cm)expC(f(i),cli) | 否 | 是 | 欧式空间 | – | cocoloss |
L-softmax | L i = − log ( e ∥ W y i ∥ ∥ x i ∥ Φ ( θ y i ) e ∥ W y i ∥ ∥ x i ∥ Φ ( θ y i ) + ∑ j ≠ y i e ∥ W y i ∥ ∥ x i ∥ cos ( θ j ) ) L_i =-\log(\frac{e^{\|W_{y_i}\|\|x_i\|\Phi(\theta_{y_i})}}{e^{\|W_{y_i}\|\|x_i\|\Phi(\theta_{y_i})}+\sum_{j \neq y_i}e^{\|W_{y_i}\|\|x_i\|\cos(\theta_j)}}) Li=−log(e∥Wyi∥∥xi∥Φ(θyi)+∑j=yie∥Wyi∥∥xi∥cos(θj)e∥Wyi∥∥xi∥Φ(θyi)) | 否 | 否 | 角度空间 | ∥ W 1 ∥ ∥ x ∥ cos ( m θ 1 ) = ∥ W 2 ∥ ∥ x ∥ cos ( θ 2 ) \|W_1\|\|x\|\cos(m\theta_1) = \|W_2\|\|x\|\cos(\theta_2) ∥W1∥∥x∥cos(mθ1)=∥W2∥∥x∥cos(θ2) | L-Softmax |
L2-softmax | L = − 1 m ∑ i = 1 m log e W y i T x i + b y i ∑ j = 1 n e W j T x i + b j L = -\frac{1}{m}\sum_{i=1}^m\log\frac{e^{W_{y_i}^Tx_i+b_{y_i}}}{\sum_{j=1}^ne^{W_j^Tx_i+b_j}} L=−m1∑i=1mlog∑j=1neWjTxi+bjeWyiTxi+byi | 否 | 是 | 欧式空间 | ( W 1 − W 2 ) x + b 1 − b 2 = 0 (W_1-W_2)x+b_1-b_2=0 (W1−W2)x+b1−b2=0 | L2-softmax |
A-softmax | L i = − log ( e ∥ x i ∥ cos ( m θ y i , i ) e ∥ x i ∥ cos ( θ y i , i ) + ∑ j ≠ y i e ∥ x i ∥ cos ( θ j , i ) ) L_i =-\log(\frac{e^{\|x_i\|\cos(m\theta_{y_i},i)}}{e^{\|x_i\|\cos(\theta_{y_i},i)}+\sum_{j \neq y_i}e^{\|x_i\|\cos(\theta_j,i)}}) Li=−log(e∥xi∥cos(θyi,i)+∑j=yie∥xi∥cos(θj,i)e∥xi∥cos(mθyi,i)) | 是 | 否 | 角度空间 | ∥ x ∥ cos ( m θ 1 ) = ∥ x ∥ cos ( θ 2 ) \|x\|\cos(m\theta_1) = \|x\|\cos(\theta_2) ∥x∥cos(mθ1)=∥x∥cos(θ2) | A-Softmax |
AM-softmax | L i = − log ( e s ( cos ( θ y i ) − m ) e s ( cos ( θ y i ) − m ) + ∑ j ≠ y i e s cos ( θ j ) ) L_i =-\log(\frac{e^{s(\cos(\theta_{y_i})-m)}}{e^{^{s(\cos(\theta_{y_i})-m)}}+\sum_{j \neq y_i}e^{s\cos(\theta_j)}}) Li=−log(es(cos(θyi)−m)+∑j=yiescos(θj)es(cos(θyi)−m)) | 是 | 是 | 余弦空间 | cos ( θ 1 ) − m = cos ( θ 2 ) \cos(\theta_1)-m = \cos(\theta_2) cos(θ1)−m=cos(θ2) | AM-Softmax |
cosFace | L i = − log ( e s ( cos ( θ y i ) − m ) e s ( cos ( θ y i ) − m ) + ∑ j ≠ y i e s cos ( θ j ) ) L_i =-\log(\frac{e^{s(\cos(\theta_{y_i})-m)}}{e^{^{s(\cos(\theta_{y_i})-m)}}+\sum_{j \neq y_i}e^{s\cos(\theta_j)}}) Li=−log(es(cos(θyi)−m)+∑j=yiescos(θj)es(cos(θyi)−m)) | 是 | 是 | 余弦空间 | cos ( θ 1 ) − m = cos ( θ 2 ) \cos(\theta_1)-m = \cos(\theta_2) cos(θ1)−m=cos(θ2) | cosFace |
ArcFace | L i = − log ( e s cos ( θ y i + m ) e s cos ( θ y i + m ) + ∑ j ≠ y i e s cos ( θ j ) ) L_i =-\log(\frac{e^{s\cos(\theta_{y_i}+m)}}{e^{^{s\cos(\theta_{y_i}+m)}}+\sum_{j \neq y_i}e^{s\cos(\theta_j)}}) Li=−log(escos(θyi+m)+∑j=yiescos(θj)escos(θyi+m)) | 是 | 是 | 角度空间 | cos ( θ 1 + m ) = cos ( θ 2 ) \cos(\theta_1+m) = \cos(\theta_2) cos(θ1+m)=cos(θ2) | arcface |
人脸识别中的损失函数总结
最新推荐文章于 2023-06-30 13:22:54 发布