Matrix Transformation-based Optimized CNN.

High-accuracy automatic recognition of facial expression is a challenging task because of the subtlety, complexity and diversity of facial expression. In this paper, we improve a traditional feature-based method, named the convolutional neural network (CNN). In order to increase the accuracy, we treat the tensor of the output layer as a multi-dimensional vector, and through matrix transformation, we can magnify its eigenvalues. Empirical evidence shows that this method makes the calculated difference between the probability values of each expression larger so that the computer obtains a smaller loss value when correctly recognized. Moreover, the training speed is faster than that before the optimization.