Analysis Of Multi-convolutional Neural Network Fusion Approach For Smile Recognition

As one of the most widely used expressions in our daily life,smile has become a research hotspot in academia and industry.It has a great potential application market,and can be used in emotional monitoring,human-computer interaction,scene analy-sis,camera shutter control and other application scenarios.However,problems such as different sizes of faces,different lighting environments,different facial postures and ex-ternal occlusion undoubtedly bring difficulties to the recognition of smiles.At present,the research progress of smile recognition methods can be divided into two categories:one is based on traditional manual feature extraction method,the other is based on in-depth learning of self-feature extraction end-to-end method.Thanks to the powerful feature learning ability of in-depth learning,it has gained more and more attention and application in the field of computer vision.This paper also proposes an optimization algorithm for laughter recognition based on deep learning method.The main contents and work of this paper are as follows:1.An optimization algorithm based on deep learning convolutional neural net-work(CNN)fusion is proposed.Before extracting features from CNN network,the algorithm first uses gray scale,histogram equalization,Gauss blur,random rotation and other means to suppress the interference of illumination,background and other factors in the image and expand the image data set.Compared with the single CNN recognition method,this method improves the optimal parameter training problem which is diffi-cult to solve in the general CNN network.It makes full use of the mouth features that change significantly when smiling as an auxiliary input source,so that it can capture more details of face images to improve the recognition rate of smiles.The simula-tion experiment on GENKI-4K database proves the effectiveness and superiority of the model fusion method.2.To overcome the shortcomings of many parameters and slow training speed of the neural network,this paper introduces the MobileNet model compression method,which uses the combination of deep convolution and point-by-point convolution instead of conventional convolution operation to lighten the network to reduce the computa-tional load and improve the operational efficiency.At the experimental stage,the time efficiency of the network model before and after compression is compared.