Research On Convolution Neural Network And Image Rocognition Technology

Image recognition technology has a wide range of applications in various fields.Such as text recognition,object recognition,traffic monitoring,photography for other violations.Research on image recognition has been going on for many years.There are two major categories in the existing algorithms.One is the traditional image recognition algorithm based on image processing.Another is a new type of image recognition algorithm based on artificial intelligence.The traditional image recognition algorithm takes a long time,but it can not achieve real-time requirements,and has low accuracy.Compared with the former,the algorithm of image recognition based on artificial intelligence is simple and fast?It is widely used because it can learn more advanced features,thereby enhancing the accuracy of image recognition.This paper focuses on acceleration and compression technology of convolution neural network and image recognition.The main purpose is to propose an image processing algorithm based on artificial neural network with low complexity,high recognition accuracy,high efficiency and fast.This article has carried on the statistical analysis to all sorts of existing artificial intelligence algorithms to chooses its corresponding optimal algorithm combination for specific problems according to the application characteristic and the insufficiency of each kind of algorithm through the experiment.Using the method of artificial neural network acceleration,we improve the algorithm of present image recognition system.The research results of this paper are as follows:1.An effective pruning method of convolutional neural network based on weight parameter is proposed,it can be easily applied to the project,and can speed up the network computation time and reduce the network model.After optimization and pruning,the accuracy of the model is less than 1%of the original,and the speed of the model is even more than ten times faster than the original.2.Based on the deep rebirth and the pruning algorithm proposed in this paper,it is proposed that the deep rebirth and pruning algorithm should be further accelerated and optimized.The optimization of the network from all aspects of network structure and model scale makes the model size and speed have more excellent performance than the single optimization.3.Aiming at the convolution neural network architecture,a new model compression method is proposed.By using large networks transformation algorithm instead of direct small network,the model memory can be optimized for compression.It can not only maintain the original large-scale network accuracy but also take advantage of the computational convenience of small networks.Learn from the mobilenet model,it turns the original network model into mobilenet network model.After optimization and transformation,the accuracy of the model is less than 1%of the original,and the speed of the model is even less than one tenth of the original.