Training Convolutional Neural Network With Only Less Positive Samples To Build Model Of Multio-bjective Recognition

A great progress has been made in the field of object detection and identification,and commercial purposes have been achieved with the continuous development of technologies both in the domestic and foreign.The recognition method based on deep learning is more accurate and robust than the traditional recognition algorithms.Recognition of all kinds of symbols has reached a high level.However,the recognition rate will drop sharply,when training samples are not sufficient because some symbols are difficult to collect in a short time.A small sample training method of convolution neural network is required to improve the reliability of the trained model and to shorten the sample preparation period.The method similar to recurrent neural networks is used to identify multiple objects to avoid the difficulty to correctly segment the object,and solve the problem of the low rate of recognition.In this paper,a convolution neural network system is proposed for training with small samples,and a method similar to recurrent neural network is used to identify the multiple objects.First of all,this paper has introduced and discussed the components;key technologies and the difficulties of preprocess before training samples of the training system.A method of transforming small samples to a large number of effective training samples is proposed in order to solve the problem that some symbols are difficult to collect.The method is based on the objects of training samples and the training samples are reasonably converted to a large number of samples.The objects in different directions are obtained by translating the object in different directions.The original samples and the processed samples are then expanded and shrinked in different scales.Dilating and thinning the training samples,different specifications of samples are obtained.Then the noises are added to the samples obtained above and a large number of effective training samples are obtained.This processing is regarded as the first part of the training system and the latter part is to train the classic convolutional neural network.The trained model has a higher recognition rates in a large number of experiments by means of adjusting and optimizing the network parameters.The system inputs a small number of samples of pictures and is finally trained to be a reliable model so as to solve the problem of lack of training samples and low recognition rates of the trained models.The idea of circular neural network is used to combine the multiple results to obtain the final results and achieve the expected results when the model is called to recognize the objects.In this paper MATLAB is used to test the recognition accuracy of the model using a large number of test samples in order to verify the effectiveness of the methods.Through a large number of experimental comparisons,the proposed method can effectively extend the small number of samples to a large number of samples and the trained model has a high recognition accuracy.All three models trained by this method for recognition of digital tube numbers,printed numbers and letters,have an accuracy of over 99%within the range of model tolerance.The Recurrent Neural Networks is used in the process of recognition of multi-objects.In the experiments correct results are obtained so that the proposed method solves the problem that multi-objects is difficult to correctly segment in the traditional method.