The Research On The Mothed Of Image Recognition Based On Convolutional Neural Networks

With the rapid development of hardware and software technology and the arrival of Big Data age,the deep learning technology have got many breakthrough results in a growing number of areas.Inspired by the biological vision system,the thought of deep learning was introduced into neural networks,thus convolutional neural networks has become one of active research direction in the filed of artificial intelligence and computer vision in recent years.With the trait of local receptive field?parameter sharing?down-sampling and sparse connection,convolutional neural networks has been widely applied in more and more different areas of social life.Based on the research of the convolutional neural networks theory and combining the feature of images,the construction method?transfer training method?optimization algorithm and the recognition performance of convolutional neural networks model on the task of image recognition are discussed in detail,the main contents are as follows:(1)A lot of methods of image recognition often divides the image recognition task into two stages: feature extraction and classification recognition.The recognition process often involves complicated image preprocessing and artificial marker feature design,considering the need to fully preserve the original feature information of the image,an end-to-end image recognition model with 11-layers based on convolutional neural networks is constructed.The recognition model can be self-learning features from the original image,and the hierarchical structure can couple the feature extraction and classification together,achieve the end to end classification and recognition.Finally,the model is applied to the license plate character recognition task under the natural scene,the contrast experiment proves the strong self-learning and efficient classification performance of this convolution model.(2)Aiming at the overfitting problem which often occurs in the training process of deep convolutional networks,this thesis construct many different convolutional networks and the parameter adjustment mothed of image data augmentation trick,adjustment of convolutional kernel size,transforming different gradient descent training optimization algorithm and adding different dropout layers with different loss rate are applied and analyzed.The contrast experiment of model training loss with different training parameters verifies the great anti-overfitting ability of the optimal recognition model.(3)To solve the problem of image recognition on few samples with insufficient training data,a transfer learning method based on convolutional neural networks is proposed.according to the difference of the data content between the source domain and the target domain and with the help of pre-trained model which has been trained completely on the large-scale image datasets,the pre-trained model in the source domain is transferred to the small-scale image datasets of the given isomorphic space and heterogeneous space;and the layer freezing method to fine-tune the network model is applied finally.The experimental results on the partial few sample public datasets and the small-scale license plate character dataset enhanced the good application performance of this proposed mothed on the small-scale datasets image recognition task.