摘 要 ：目的 为解决真实环境中由类内差距引起的面部表情识别识别率低，针对室内外复杂环境对类内差距较大的面部表情识别难度大等问题。方法 提出一种利用生成对抗网络GAN识别面部表情的方法，在GAN生成对抗的思想下，构建一种IC-GAN(Intral-Class Gap GAN)网络结构，使用卷积组建编码器、解码器对自制混合表情图像进行更深层次的特征提取，使用基于动量的Adam优化算法进行网络权重更新，重点针对真实环境面部表情识别过程中的类内差距较大的表情进行识别，使其更好的适应类内差异较大的任务。结果 本文基于pytorch环境在自制的面部表情数据集上进行训练，在面部表情验证集上进行测试，并与DBN深度置信网络、GoogleNet网络做对比实验，最终IC-GAN网络的识别结果比DBN网络提高11%，比GoogleNet网络提高8.3%。结论 经实验验证了IC-GAN在类内差距较大的面部表情识别中的精度，降低面部表情在类内差距较大情况下的误识率，提高系统鲁棒性，为面部表情的生成工作打下坚实的基础。
Depth learning recognition method for in-class gap expression
chenliang,wupan,liuyunting(Shenyang University of Technology)
Objective Complex environmental safety issues such as subways, railway stations, and airports are increasing, and unstable events are frequent. The factors affecting facial expression recognition in security screening are large. The larger intra-class gap seriously affects the accuracy of facial expression recognition. In order to solve the problem of large gaps in facial expression recognition in the real environment, it is possible to identify the suspicious molecules being monitored, and let the security personnel prepare in advance to accurately identify them. Facial expressions are especially important for preventing safety issues.Complex environmental safety issues such as subways, railway stations, and airports are increasing, and unstable events are frequent. The factors affecting facial expression recognition in security screening are large. The larger intra-class gap seriously affects the accuracy of facial expression recognition. In order to solve the problem of large gaps in facial expression recognition in the real environment, it is possible to identify the suspicious molecules being monitored, and let the security personnel prepare in advance to accurately identify them. Gwen Littlewort et al. used the Gabor transform to extract the expression features of the upper half face, the lower half face and the entire face respectively. Two SVMs (Support VectorMachines) were used to construct 21 SVMs for the 7 basic expressions (one-to-one method), and then The classification results of each classifier are merged by nearest neighbor method, voting decision method and MLR (multinomial logistic rldgeregression) method, and classified by two classification methods. Experiments show that the fusion method using MLR is the best, achieving a recognition rate of 91.5%. Hal Hong et al. established a corresponding expression database (7 expressions, 28 images each with different expression intensity) for different faces as training samples, and used the elastic map matching method for training recognition. The era of big data has arrived. With the advancement of computer hardware, deep learning continues to advance. Traditional facial facial expression recognition methods can no longer meet the needs of the development of the times. New algorithms for facial expression recognition are also coming, based on depth. The method of learning is more widely used in facial expression recognition.Facial expressions are especially important for preventing safety issues. Although facial recognition intelligent recognition technology has a long history of research, a large number of research methods have been proposed, but due to the large gap in facial expressions, complex expressions and many influencing factors, the current intelligent recognition of facial expression results is unsatisfactory. Deep learning, due to its powerful expressive ability, has just begun its research in the field of real-life facial expression intelligent recognition. In the next period of time, considerable progress will be made in the real-world facial expression recognition field. This paper further studies the real-life facial expression recognition based on deep learning. Method A method for identifying facial expressions using a generated confrontation network GAN is proposed. The IC-GAN (Intral-Class Gap GAN) network structure is constructed for the difficult security of complex environments such as subways, railway stations, and airports. The network uses the convolutional layer, the fully connected layer, the active layer, and the BN (BathNorm) layer. The Softmax layer is composed of a convolutional encoder and a decoder for deeper feature extraction of facial expression images to ensure the accuracy of facial expression recognition, and the use of mixed facial expression data based on real environment based on downloading and parsing video from the network. Set, the image is expanded, and the facial expression data is normalized to better adapt to the task of large difference within the facial expression recognition. This paper is based on pytorch to train facial expression data. Network training and network recognition use momentum-based Adam to update network weights, adjust network parameters and optimize network structure based on this, test on facial expression verification set, and DBN Deep Confidence Network and GoogleNet conduct comparative experiments, increasing the complexity of facial expression features with a large number of intra-class differences, and improving the accuracy of facial expression recognition. Result Compared with other traditional methods and the recognition of DBN, GoogleNet and other networks, facial expression recognition based on deep learning is more accurate. The experimental results show that the IC-GAN recognition model network is more accurate. When the input image is 256×256, the recognition result is more accurate. The DBN network is 11% higher and 8.3% higher than the GoogleNet network. Conclusion While ensuring the accuracy, the IC-GAN network model reduces the misrecognition rate of expressions in the case of large gaps in the class, blurred images and incomplete facial expressions, and improves the robustness of the system. To lay a solid foundation for the generation of facial expressions.