Design Of Lightweight Face And Expression Recognition Model And Its Application On Mobile Terminals

As one of the most typical application scenarios of deep learning,face recognition technology has developed rapidly in recent years.Then,as a result,the potential of facial expression recognition in health monitoring,psychological analysis,human-computer interaction and other fields was gradually explored,and nowadays it has become an emerging research hotspot.On the other hand,because of the continuous improvement of mobile artificial intelligence application requirements and the great enrichment of edge computing theory,the deployment of neural networks on mobile devices were more important than ever.But now,the application of facial expression recognition on the mobile terminal is still in its infancy,and is still facing challenges such as complex model structure and poor operational stability and so on.We adopted a multi-level optimization strategy,designed a lightweight and high-precision face and expression recognition model in this paper,and realized the deployment and application on the Android side.The main research results and innovations obtained in this paper include:(1)We constructed a lightweight face detection network Model_Face based on MTCNN by introducing structural optimization strategies such as depthwise separable convolution.The model reduced the total parameters and floating-point calculations at all levels by 22.3%and 70%respectively,and the accuracy on the FDDB dataset was only reduced by 0.84%.We used coarse-grained pruning and quantization retraining to furtherly process the face model,and converted it to a fixed-point model stored in INT8.The amount of parameters was compressed by 20%during the procedure,the storage usage was reduced by more than 60%,while the detection accuracy only dropped by 2.68%.(2)In accordance with the construction idea of low-parameter networks,we designed a lightweight facial expression recognition network Model_Emotion,which only contained 54071 total parameters,achieved 65.37%validation set accuracy on Fer-2013,and effectively avoided overfitting.We quantized and retrained Model_Emotion and converted it to INT8 format,and furtherly reduced the storage requirement of the model by 72.8%,while the detection accuracy only dropped by 0.82%.(3)The Android-side expression recognition application was finished and tested after the models were done.The recognition accuracy rate of the system for six common expressions,including smile,anger,surprise,sadness,disgust and fear,was above 80%,and the detection speed reached 50 FPS,which could still remain stable under camera movement or low light conditions.