A Class Of Neural Network Parameter Adaptive Update And Image Classification

Image classification is one of the important topics in the field of image research at present and it is necessary to find a method that can classify different types or changed image data sets with universality and efficiency.Image classification methods commonly include manual feature extraction methods,semantic-based classification methods and neural network-based classification methods.The classification method based on neural network is more popular at present.Because this method can learn images and extract image features through self-learning,it has a strong ability to recognize,classify images and it has a fast calculation speed.However,the classification method based on neural networks also has disadvantages.For example,when the training sample changes,the trained neural network model is not suitable for the training sample that has changed.Retraining with a neural network model again requires a large number of new samples.In order to avoid the problems of neural network retraining,classification accuracy reduction and a large amount of time consumption,this paper will propose a new method for image classification based on adaptive update parameters of neural networks.The main innovations are as follows:(1)A performance evaluation method of network classification model is proposed in order to solve the problem of different classification accuracy when different neural network models classify the same data set.Firstly,an evaluation criterion is established in order to find a neural network model that is most suitable for a certain type of data set for classification.Secondly,the images are extracted from the features and different network models are used for the same data set,then perform training and classification in order to find the network model structure that is most suitable for the type of data set to be tested.Finally,different network models are used to evaluate and analyze the classification results of this type of data set.(2)A parameter adaptive update method in a network classification model is proposed in order to solve the problem that the changed data set is no longer suitable for classifying it using the previously trained network model.Firstly,the convolutional neural network model parameters are used for training.Secondly,an adaptive update method of network parameters based on Kalman filtering isestablished to avoid retraining of network parameters in combination with new samples when new samples are obtained under illumination drift.Finally,the effectiveness of this method is verified by experimental simulation results.(3)A neural network model parameter adjustment method is proposed for high-noise images in order to solve the problem that the image suffers from high pollution(such as high noise or distortion)and the pixel information of part of the image is lost when the parameter adaptive update method in the simple network classification model is not suitable for such image classification.Firstly,the parameters of each dimension are extracted because the information of the increased highly polluted sample image is incomplete and each feature dimension in the image with missing information has different effects on the network output.Secondly,the average influence function is established on this layer in order to determine the weight value of each variable and the parameters are adaptively updated again using Kalman filtering.Finally,the effectiveness of the method is verified by the effect of image classification.