Image Recognition Method Research Based On Convolution Neural Network

With the improvement of computer ability and the development of science and technology,the application of image recognition technology is more and more extensive.The traditional method of image recognition is to preprocess the image at first,and then extract the features manually by a large number of experiments and specialized domain knowledge.The recognition performance often can’t satisfy the need in the face of large number of categories and complex environments.In recent years,image recognition technology based on deep learning has become a research hotspot in the field of artificial intelligence,image recognition method of deep learning compared to the traditional without complex image preprocessing process,and the network can learn features without manual designed features.Convolution neural network is a model of deep learning and it is the most widely used in image recognition,this paper studied the image recognition method based on convolution neural network,improved the convolution neural network based on the classical model LeNet-5,and applied the improved network model to oil pipeline steel recognition,the work of this paper is as follows:First of all,In order to obtain the diversity characteristics of images,the paper proposed the convolution neural network with improved structure,the structure introduced the recursive neural network based on classical convolution neural network structure,the images were put into the first convolution neural network to extract the characteristics of shallow,and then through the parallel level second network composed by the recursive neural network and convolution neural network to study the deep features,finally put the fusion features learned to classification.The experiment results showed that the improved convolution neural network structure is more advantageous to obtain the image diversity characteristics,and under the same experimental conditions,the recognition rate is higher than the traditional network.Secondly,the paper proposed the convolution neural network algorithm based on improved Fisher criterion.The algorithm improved the cost function of the convolution neural network and introduced the improved weighted Fisher criterion in the least square error cost function.The algorithm ensured the minimum distance between the actual output of the system and the expected output,and at the same time ensured the distance between similar data is closer and the distance between heterogeneous data is farther.Experimental results showed that the proposed algorithm improves the recognition rate,especially when thenumber of iteration or training data is less,the improve of the recognition rate of this algorithm is more obvious.Finally,the improved convolution neural network structure was applied to the oil pipeline steel recognition,the paper transformed the initial image into gray image,and then used the Canny operator to detect edge of the image,for the tilt image the paper corrected the tilt image by the Hough algorithm,at last divided the steel image into images contained only one character,finally constructed the training set of steel image.The paper studied the training problem of convolution neural network of small sample set,solved the over fitting problem of training process by expanding the sample and transfer learning.The experimental results showed that the results of the method proposed in this paper is better than the traditional method.