| Convolutional neural network can autonomously learn features extracted from images through neural networks,and has the advantages of local response,weight sharing,etc.It has been widely used in facial expression recognition.However,the structure of shallow convolutional neural networks can not extract facial expression features with strong expressiveness,thus affecting the effect of expression recognition.At the same time,the traditional pooling algorithm still has the disadvantages of lack of flexibility and single feature extraction.Therefore,this dissertation focuses on the feature extraction part and pooling algorithm in convolutional neural network,which is of great significance for improving the accuracy of expression recognition.This thesis proposes a convolutional neural network algorithm combined with local binary pattern(LBP)algorithm for face expression recognition.This algorithm extracts the LBP features and uses the LBP feature matrix as a convolutional neural network before extracting features from a convolutional neural network.With the input,the facial expression texture information contained in the features extracted by the convolutional neural network is enhanced,and interference factors such as illumination and background in the image are suppressed.Therefore,the expressiveness of the facial expression feature extracted by the algorithm is also fully improved.The experimental results show that compared with the original face convolutional neural network facial expression recognition method,this algorithm can effectively improve the accuracy of expression recognition and improve the concussion phenomenon of the loss function convergence curve in the convolution neural network training process.This thesis also proposes an improved dynamic adaptive pooling algorithm based on the characteristics of self-adjusting parameters of back-propagation(BP)algorithm in convolutional neural network method.The algorithm is calculated by BP algorithm according to the loss function in the training process.The gradient of the pooling parameters is continuously updated by the gradient descent method until the pooling parameters are different.The different pooling region parameters are all the same and have a certain degree of flexibility,which can overcome the single extraction feature and lack of flexibility in the traditional pooling algorithm.Insufficient,the simulation experiment on the CK+ dataset verifies the feasibility and superiority of the improved dynamic adaptive pooling algorithm. |
