Research Of Anti-noise And Energy-efficient Deep Neural Networks For Spontaneous Facial Expression Recognition

Deep Convolutional Neural Network(CNN),inspired by the research on visual cortical cells in cats,was proposed on the basis of the concept of receptive field,and has been widely used in computer vision tasks,such as facial expression recognition.Deep CNN has achieved a high accuracy rate in the recognition of "posing" facial expressions,but the recognition effect of spontaneous facial expressions in real scenes is not ideal.The main reason is that the expression images in real scenes are often noisy due to lighting conditions,shooting equipment and other reasons.On the other hand,the facial expression recognition process based on deep CNN has high power consumption,which limits its application of spontaneous facial expression recognition on mobile terminals or embedded devices.For this reason,we studied the output response characteristics of biological neurons,and Improved the Noisy Softplus(NSP)activation function which could characterize the output response characteristics of the neuron model Leaky Integrate and Fire(LIF).The INSP(Improved Noisy Softplus)was proposed to implement a deep Residual Network with strong noise resistance.At the same time,we studied the construction of deep Spiking Neural network(SNN)based on deep CNN training and transformation,and a low-power deep SNN with INSP for training and transformation of VGG-19 was proposed and implemented.The main research contents of this paper are as follows:Study on Deep residual networks with noise robustness.Firstly,to solve the problem that the error between NSP function and LIF neuron response characteristic increases layer by layer when training deep CNN,a scale factor S is introduced to adjust the release rate layer by layer,and the INSP function is obtained.Then,a deep Residual Network is built based on deep CNN,and the network structure is adjusted to reduce the computational complexity.Finally,the improved NSP is applied to the deep Residual Network as the activation function to solve its different implementation problems inside and outside the residual structure,and the deep residual network with noise robustness is obtained.Study on Energy-efficient deep SNN construction based on Deep Convolutional Neural Network training and transformation.Firstly,we analyzes the reasons for the accuracy loss when CNN is converted into SNN and gives a general solution.Then,it proposes and implements the application of INSP to train the cropped vgg-19,constructs the corresponding deep SNN method,and makes a comparison with the existing method combining training with weight adjustment with the application of ReLU.Finally,the power consumption of the constructed deep SNN and VGG-19 is estimated.The proposed algorithm is verified experimentally in the spontaneous facial expression database.Firstly,aiming at the shortcomings of the small sample size of spontaneous facial expression,the sample data were enhanced to prevent overfitting.Then,according to the improved algorithm,a comparative experiment is carried out on spontaneous facial expression database etc,which verifies the feasibility and advancement of this method from the aspects of recognition performance and anti-noise performance.