| With the rapid development of information technology and high-tech warfare technology,air supremacy plays an extremely important role in modern warfare.The planes can not be restricted by obstacles on the ground,has the ability to fly at high speed in three-dimensional space,and can generate tremendous destructive power.It is an important offensive weapon used to compete for air supremacy in modern warfare.In order to be able to accurately and effectively attack the enemy’s planes militarily,it is required that its own weapons be able to cope with various complex battlefield environments and diverse climate changes,possess anti-stealth ability and strong anti-interference ability,have high recognition rate and high guidance accuracy.This subject has been established based on the fact that infrared imaging technology has a long-range effect,strong anti-interference ability,high measurement accuracy,and the dominant position of deep learning in the field of image recognition,aiming to achieve real-time detection of plane targets in infrared images through convolutional neural network technology.This work takes the plane targets in infrared image as the research object,and uses the convolutional neural network method to realize the real-time detection of the plane targets in the infrared image.The specific research content is as follows:(1)For the research object of this paper,the production of the infrared plane target dataset was completed.The author collect the infrared plane images containing different scenes,types,poses and size from the Internet as materials of the dataset.Considering that the number of infrared datasets is relatively small and it is easy to cause overfitting during training,this paper uses six image augmentation methods,such as horizontal tumbling,rotation,brightness transformation,shearing,zooming in and out,and adding Gaussian white noise to expand the infrared plane image dataset.(2)By applying the RCNN(Region with Convolution Neural Networks features)target detection framework to the infrared plane target detection field,this paper verifies the feasibility of using convolutional neural networks to solve such problems.Considering that the traditional method of region of interest extraction in RCNN model will bring huge amount of computation and generate a large number of redundant frames,this paper proposes and implements a new method of region of interest extraction based on the characteristics of infrared images.The grayscale saliency of the image achieves segmentation of the plane’s target and background,thereby greatly reducing the computational time.In order to reduce the computational complexity of the network and to avoid overfitting due to the lack of model generalization capability,a simple 6-layer neural network was designed to extract image features.By training the network models and feature classifiers in the RCNN model framework,the trained model framework is used for results testing,which verifies the feasibility of deep learning methods to solve such problems.(3)For the shortcomings of long time frame,incapable of end-to-end training and inaccurate positioning of RCNN model framework,this paper also uses the SSD model framework to achieve the detection of infrared planes.In this paper,the original VGG network is improved,and a complete SSD(Single Shot Multi Box Detector)model framework based on the improved VGG16 main body network is given.By improving the cross-entropy loss function,the value of the loss function when predicting correct positive class samples is reduced,thereby optimizing the class loss function and making the model more focused on difficult misclassified samples.The class prediction loss function is optimized.The SSD network is trained through fine-tuning migration learning,and the experimental results are compared with the RCNN model framework in terms of speed and detection rate.The test has achieved real-time detection of the plane targets in infrared images on the basis of ensuring the detection accuracy. |
PDF Full Text Request