Target Recognition Based On Deep Learning

The key issue of target recognition is how to obtain a good description of the data,namely feature extraction.Traditional feature extraction methods are faced with many difficulties and limitations in the complex big data environment.On the one hand,traditional feature extraction methods require researchers to have professional background knowledge and data processing skills,on the other hand,the steps of traditional feature extraction are complex.However,deep learning can automatically learn the representation of features from big data,designers are not required to have strong prior knowledge.Meanwhile,traditional data preprocessing,feature extraction,classifier design and other steps are adopted in an end-to-end adaptive optimization process,which is convenient for the deployment and application of deep network tasks.Based on the deep learning theory,the main research contents and achievements of target recognition task are as follows:1.An image recognition method based on feature fusion of deep multi-perceptual interest regions is designed.By visualizing the heat map and character visualization of a single classification model,the method obtains conclusions that the different feature regions associated with different network models.On this basis,multiple model fusion mechanisms are designed,and the characteristics of different models are merged to obtain the fusion network model MVF(Multi View Fusion),MVF-tiny(Multi View Fusion-tiny)and a new network model fusion method Voted Model.From the experimental results,it can be concluded that this method significantly improves the classification accuracy compared with the single model in the test data,which proves the effectiveness of the fusion method in this paper.2.A signal coding modulation recognition method based on one-dimensional convolution modeling is designed.For the characteristics of one-dimensional convolution weight sharing and multi-scale sensing field,and aims at the problem of signal fine granularity classification,a one-dimensional convolution model is designed for the classification of communication coded modulation signals.Experimental results show that the degree of discrimination between different modulation modes is obvious,but different frequency signals of the same modulation mode are easy to form a wrong conclusion.Finally,in order to meet the speed of radio classification and the requirements of embedded system for model deployment,this paper designed a one-dimensional light weight network to reduce model operation time and obtain high accuracy.3.A signal coding modulation recognition method based on deep learning time series modeling is designed.In order to obtain the sequence correlation between signals,the time-sequence neural network is used to extracts the characteristics of the signals.Firstly,feature extraction is carried out through one-dimensional convolution,aiming at reducing the dimension of parameter space.Secondly,the temporal features of signals are mined by constructing a temporal network.The experimental results show that the network model designed by this method has the same accuracy rate as the one-dimensional convolution model,and the time-series network model constructed by this method is effective for signal coding and modulation recognition.