基于角度特征的分类网络

Classification Net Based on Angular Feature

卷积神经网络(Convolutional Neural Networks,CNN)在图像分类任务中的卓越表现,使得其被广泛应用于计算机视觉的各个领域。图像分类模型精度与效率的提升,除了归功于网络结构的改变外,还有很大一部分原因来自于归一化技术以及分类损失函数的改进。在人脸识别任务中,随着精度的不断提升,分类损失函数从Softmax Loss到Triplet Loss,又从L-Softmax Loss到Arcface Loss,度量方式从几何度量发展到角度度量。度量方式的改变实际上是特征形式的变化,即特征形式从一般特征转变为角度特征。在Mnist数据集上,使用角度度量损失函数训练得到的特征点呈角度分布,同时准确率比几何度量高;将角度度量方式用更直接的角度特征来表示,训练得到的同类特征点呈直线分布,准确度也比一般角度度量更高。这不禁令人思考,在CNN分类模型中是否可以使用角度特征来代替一般特征。在CNN分类模型中,其主要架构往往由多个卷积层和一个或多个全连接层组成,通过统一卷积层与全连接层的归一化操作,得到角度卷积层与角度全连接层。在普通分类网络的基础上,用角度卷积层替换卷积层,用角度全连接层替换全连接层,可以得到一个由角度特征组成的角度分类网络。在Cifar-100数据集上,基于ResNet-32构造的角度分类网络相比原分类网络,分类准确率提高了2%,从而论证了角度特征在分类网络中的有效性。

The excellent performance of Convolutional Neural Networks(CNN)in image classification tasks makes CNN models widely used in various fields of computer vision.In addition to the changes in the network structure,a large part of the reason why the accuracy and efficiency of the image classification model increase year by year comes from the normalization technology and the improvement of the classification loss function.In the face recognition task,with the increasing precision,the classification loss function change from Softmax Loss to Triplet Loss,and from L-Softmax Loss to Arcface Loss,the measurement method develops from geometric measurement to angle measurement.The change of measurement mode is actually a change of feature form,and the feature form changes from general feature to angle feature.The feature points trained on the Mnist dataset using the angle metric loss function are angularly distributed,and the accuracy is higher than the geometric metric.If the angle metric is represented by more direct angular features,the feature points of the same class are linearly distributed after training,and accuracy is also higher than the general angle metric.This makes people wonder whether angle features can be used instead of general features in the CNN classification model.In the CNN classification model,the main structure is often composed of multiple convolutional layers and one or several fully connected layers.Through unifying the normalization operation of the convolutional layer and the fully connected layer,layers in model come to the angular convolutional layers and the angular fully connected layers.On the basis of the common classification network,the convolution layer is replaced by the angle convolution layer,and the full connection layer is replaced by the angle full connection layer,and then an angle classification network composed of angular features can be obtained.The accuracy of the angle classification network constructed on ResNet-32 is 2%higher than that of the original classification network on the Cifar-100 dataset.The validity of the feature in the classification network is demonstrated.