Research On Image Classification Algorithm Based On Circular Convolutional Neural Network

With the development of artificial intelligence,target detection and classification based on image become the basis of other high-level visual tasks such as target tracking,individual identification and event analysis.Aiming at the problem of image objects' detection and classification and the visual recognition ability which Convolutional Neural Network model after that of human being,the significant feature of the extracted image object which is used for image object detection and classification shows good results.About the traditional CNN model,the more the number of convolution samples,the more obvious the effect of the final target detection and classification is.However,with the increase of the depth of CNN,the difficulty of training parameters is doubled.The calculated amount is greatly improved,so that the depth of the Convolutional Neural Network can not be applied on some devices with high real-time requirements and slower computation speed.At the same time,the training of DCNN is more difficult,so the parameter training is a challenge which the traditional CNN is facing.The Batch stochastic gradient descent algorithm is a popular algorithm in the process of training neural networks.In the training process of the CNN by using the Batch stochastic gradient descent algorithm,the gradient will drop rapidly when the error rate and the loss value are large,but In the case of Neurons saturated,the rate of gradient decline will slow down,the training neural network will appear over-fitting,the image target detection,the rate of false recognition and the loss value of cost function will appear random small fluctuations,so that the neural network model training will becomes difficult,secondly,the Learning Rate of the stochastic gradient descent algorithm needs to be set manually.In the training process of real neural network,the Learning Rate is too large to cause the training process to be difficult to converge and the global optimal solution can not be found.If the learning rate is too small and the speed of gradient descent is too slow,Local optimization will be difficult to find the global optimal solution.On the other hand,the rate of gradient descent will be slow when the stochastic gradient descent algorithm is in the face of the non-convex optimization problem in the training process.Therefore,the training speed of stochastic gradient descent algorithm is slow in the process of training the CNN,and sometimes can not find the global optimal solution.This paper is to realize the detection of target and classification of the image.Firstly,the Convolutional Neural Network(CNN)and the Recurrent Neural Network(RNN)are used to construct the cyclic convolution neural network,with the Trial retreat algorithm and cyclic convolution neural network model trained by normalized parameter and learning rate from the Batch stochastic gradient descent algorithm based on Golden Section Algorithm.After that,connections among the neural network are increased.Deep cyclic convolution neural network model(RCNN)is used to extract the image target characteristics and Softmax regression classifier is used to sort,to achieve the target detection of image and classification.Finally,train the cyclic convolutional neural network(RCNN)model with the improved BFGS algorithm,and extract the characteristics of the image for the input layer of the limit learning machine with trained cyclic convolution neural network,and realize the target detection of image and classification.Based on MNIST handwritten digital character data set and CIFAR-10 data set to test the target detection of image and classification,Its results show that it can achieve better detection classification and faster convergence in the detection of target and process of classification,by using cyclic convolution neural network model trained by the Trial retreat algorithm and the Batch stochastic gradient descent algorithm based on Golden Section Algorithm and cyclic convolution neural network trained by the improved BFGS for the input layer of the limit learning machine.