Research On Target Recognition Method Based On Machine Learning

Target recognition has always been a research hotspot in the field of computer vision.Target recognition is widely used on the battlefield.However,due to the complex battlefield environment,there are problems in the target recognition process,such as inaccurate feature selection information and low recognition rate of single classifier..In thesis,an end-to-end deep convolutional neural network method is proposed to identify the target.Firstly,semi-supervised enhanced migration learning is proposed for the stability of weight distribution of deep convolutional neural networks,data dependence problems and the gap between source and target domains.The deep convolutional neural network is not easy to converge in the early stage,and the weight distribution is easily affected by the data distribution.The introduction of semi-supervised migration learning can solve such problems to some extent.For the semi-supervised migration learning,the over-fitting problem is proposed.The enhanced migration learning is proposed,and the over-fitting phenomenon is mitigated by fusing multiple classifier parameters.Experiments on the Cifar-10 dataset show that network training with semi-supervised enhanced migration learning is more stable and converges faster.Then,the reuse feature selection mechanism is proposed for the problem that the attention mechanism has a large amount of full connection parameters and the distribution truncation problem caused by the grouping.Selecting better features can greatly improve the accuracy of the model,and the attentional mechanism can enhance the salient features in the deep neural network.In thesis,the convolution feature selection is proposed,and the convolution calculation is used to replace the full-join operation in the weight mapping phase of the attention mechanism.The feature map can be better expressed in the case of reducing the parameter amount.Aiming at the problem of distribution truncation of feature map group selection,thesis proposes multiplexed slice grouping and combined parallel selected features for feature selection,which reduces the information loss and over-selection problems caused by truncation of feature graph distribution and increases the recognition rate.Experiments on the STL-10 dataset show that the proposed reuse feature selection network(CoFS-Net)is better than other target recognition networks.Finally,combined with the methods proposed in the third chapter and the fourth chapter,the practical application scheme of the machine learning target recognition method in the battlefield is proposed.The reuse feature selection network proposed in Chapter 4 is used as the feature extractor.Supervise enhanced migration learning to initialize parameters.In thesis,the target recognition and recognition task is divided into two steps.Firstly,the second classifier is used to judge whether the target exists,and then the multi-classifier is used for target classification.Experiments were carried out on the collected battlefield target datasets,and experiments showed that the proposed method is better than other methods.