一种基于自校准卷积残差网络的年龄识别方法

An Age Recognition Method Based on Self-calibrated Convolutional Residual Network

传统卷积神经网络(CNN)提取人脸面部年龄特征信息时受限于感受野,易导致年龄识别准确率较低,本文建立了一种基于自校准卷积残差网络(SC-ResNet)的年龄识别方法。首先对输入图片进行裁剪和归一化预处理;然后在残差网络中用3×3卷积提取局部表观特征,再通过自校准SC-block模块进一步扩大局部特征提取范围,并将两者校准融合,获得更丰富的面部语义信息;最后采用Softmax结合交叉熵损失函数进行更精细化的年龄分类;实验中网络参数采用自适应矩估计(Adam)算法优化。结果表明,本文SC-ResNet模型能更好提取年龄特征,年龄识别率达到96.3%,比原始的残差网络高出5.8个百分点。

Traditional convolutional neural network(CNN)is limited to the receptive field when extracting facial age feature information,which leads to low accuracy of age recognition.In this paper,an age recognition method based on self calibration convolutional residual network(SC-ResNet)is proposed.Firstly,the input image is preprocessed by clipping and normalization;then the local apparent features are extracted by 3*3 convolution in residual network;then the local feature extraction range is further expanded by self calibration SC-block module,and the two are calibrated and fused to obtain richer facial semantic information;finally,softmax combined with cross entropy loss function is used for more refined age classification The network parameters are optimized by adaptive moment estimation(Adam)algorithm.The results show that the SC-ResNet model can extract age features better,and the age recognition rate reaches 96.3%,which is 5.8 percentage points higher than the original residual network.