Research On Algorithm Of Image Recognition System Based On Deep Learning

With the rapid improvement of computer hardware performance,the development of massive data learning algorithms and the in-depth study of neural networks,deep learning methods are widely used in image recognition tasks.Convolutional neural networks are especially outstanding in image recognition,and show excellent performance in such fields as speech recognition,motion prediction,medical diagnosis and so on.Compared with traditional artificial neural networks,convolution neural networks simulates the response of visual neurons to image information better.The sparse connection method greatly reduces the scale of the parameters of the neural network and the process of extracting image features repeated by each convolution kernel of the layer minimizes the number of free parameters and effectively improves the training efficiency of the model.The image recognition method based on convolution neural network avoids the complicated and inefficient feature extraction of traditional methods,and integrates image feature analysis into the process of model training.Although convolutional neural network models have made remarkable achievements in image recognition tasks,in the face of mobile devices,embedded systems and other platforms with limited computing resources,the traditional model based on deep learning is no longer an ideal choice because of its large amount of parameters and computation.Therefore,under the premise of guaranteeing the performance of the model,this paper makes an in-depth study on how to simplify the network structure and parameters of the model,and proposes an improved model based on the original MobileNet.The main innovative research work of this paper mainly includes the following three points:1.In order to improve the performance of existing models,deepen the layers o f model structure and increase the number of parameters,MobileNet is choosen as the basic research model after reading a lot of literature in this paper.The innovation of this model is to replace the traditional convolution operation with the deep separable convolution operation,which is the combination of two-dimensional deep convolution and three-dimensional point-by-point convolution.The traditional method of convolution multiplication is changed to the combination of multiplication and addition,which greatly reduces the amount of parameters and computation of the model,while maintaining the accuracy requirements of image recognition tasks.On the basis of the original model,the average pooling layer of the model has been improved.The performance of image salient features will be reduced by the application of the average pooling layer.The method of global depth convolution is used to optimize and improve the problem in this paper.2.This paper makes a deep research on the popular optimizers in deep neural networks,summarizes and analyses the advantages and disadvantages of each optimizer,and proposes that RMSprop optimizer used in original MobileNet should be replaced by Adam optimizer with better comprehensive performance.Through the experimental verification of this paper,The stability and accuracy of training model can be improved by this method.3.In view of the long time-consuming problem of model training,the transfer learning method is used to apply the trained model parameters on the ImageNet data set to the experimental model in this paper,which effectively improves the convergence speed and accuracy of the model.Combining the above three methods to improve the original model,the proposed model has improved the training accuracy by abo ut 1% and the test accuracy by about 3% compared with the original model in the dataset used in this paper.The over-fitting rate is also better than the original model.