Research And Applications Based The Improved Convolutional Neural Network

Convolutioanl neural network is one of the new kinds of neural network, which is combined with artificial neural network and deep learning technology. It is the first algorithm model which can train multi-layer network structure successfully. Convolutional neural network is a multilayer perceptron designed to recognize 2-D images specially, which has good self-learning ability, fault-tolerant ability, and has high degree of invariance to translation, scaling, inclination or other forms of deformation. Based on the traditional convolutioanl neural network, this paper proposes three improved algorithms:1. We propose a convolutional neural network algorithm model based on the weighted Fisher criterion. The algorithm mainly introduces weighted Fisher criterion to improve the cost function based on the minimum square error function. Its main purpose is to get the smallest residuals between the practical output and sample labels of the images. Meanwhile, it makes the samples in small distance with-class and large distance between-class.2. We propose a convolutional neural network algorithm model based on the improved activation function. The algorithm mainly combined with the advantages between two common activation functions which are named ReLUs fuction and Softplus function. The improved activation function not only can have the sparsity of ReLUs fuction, but also possess the smoothness of Softplus function. Considering the problem of imformation loss, we propose and compare two kinds of structures.3. We propose a convolutional neural network algorithm model based on the improved Gabor filter. Firstly, we improve the traditional Gabor filter and the improved Gabor filter not only can be extracte the multi-direction and multi-scale information in the image, but also has a very strong expression ability to bend at the edge of the image. Secondly, we use the improved Gabor filter convolve with the input images to get multiple directional features as the new input of the convolutional neural network instead of the original images.In this paper, the experimental verification is carried out in the Mnist handwritten database, AR face database and ORL face database on multiple data sets. The final experimental results show the feasibility of this innovative work.