A Training Method With Pixel Random Migration Expanding Small Data Set

With the deepening of the research in the field of image recognition by experts and scholars at home and abroad,a great progress has been made in image multi-classification tasks,multi-target recognition in images and image segmentation technology.There are Gabor feature,SIFT operator,LBP operator in the traditional field and Res Net network model,Goole Net network model in deep-learning field.Some of these algorithms have been applied to the daily life.As a result of the popularity of the network,huge amounts of data can be easily obtained.Therefore,the performance of deep-learning recognition method which requires a large amount of data is getting better and better.The accuracy of this method is much higher than that of traditional recognition algorithm and this method has better robustness and generalization performance.However,the following problems still exist in the training of the deep neural networks:When the data sets are short of samples or the sample distribution for each class is unbalanced,the samples cannot cover the whole feature space.This will lead to a sharp decline in the accuracy of deep learning recognition algorithm.Under these circumstances,a training method using small-scale data sets is needed to improve the accuracy and robustness of the model.Accordingly,this dissertation proposes a method of training neural network using small-scale data sets and random pixel migration expansion method based on Gauss kernel function.Combining with the commonly used method of sample expansion,it is reasonable to expand the sample to the size of the data which is necessary to complete the training.This method solves the problem in the training of a deep neural network when the number of samples is insufficient.The main work of this dissertation is as follows:(1)Aiming at the problem that the number of samples is insufficient to complete the training of neural network,a random pixel migration expansion method based on Gauss kernel function is proposed.Based on the normal distribution of the Gauss kernel function,the location of the pixels is randomly transferred.In this way,the sample expansion is completed.Sample features are transformed and reconstructed to a certain extent,which can cover more complete feature space more effectively than conventional methods.In training,more diverse features could be learned to increase the robustness and generalization ability of the model.(2)MNIST and CTW data sets are used to verify the algorithm: On the premise of guaranteeing the balance of the number of samples in each category,the model is trained.Then the training results of the original training data set and the expanded data set are compared and analyzed.The experimental results show that the convergence period,accuracy and false detection rate of the network trained with only a small number of samples are similar to those of the network trained with a large number of original data sets.It is indicated that the extended method is effective in simulating the original samples.(3)A candidate region extraction method for small target is proposed.A small target recognition system based on random pixel migration method is constructed and applied to production and life such as fabric defect detection.After collecting a certain number of original fabric defect samples,the sample set is expanded by the means of pixel random migration method.The expanded samples and the original samples are input into the network structure to train the model,and the test results of the model are counted.Statistical and comparative data show that the extended data set method in this dissertation improves the accuracy and reduces the false detection rate significantly,and has better robustness.