Research On Deep Learning Based Object Detection Algorithm In Complex Eletromagnetic Environment

In the fields of signal processing and radar imaging,Synthetic Aperture Radar(SAR),as a new technology of radar imaging,plays an important role.It is a widely used research task in the field of image processing to generate SAR images and then detect objects.Since the SAR image is generated by a band of electromagnetic wave emitted by the radar,the scattering characteristics of the image target to the electromagnetic wave vary greatly and the size changes drastically.In the imaging process,speckle noise will inevitably be generated.At the same time,a large area of background coverage will contain confusing geographical elements,which seriously affects the detection performance.In the complex electromagnetic environment,due to the presence of noise clutter and complex geographical elements,it is very difficult to visually identify the objects to be detected in SAR images.Therefore,it is of great significance to study the SAR image denoising algorithm and the SAR image object detection algorithm which is more efficient,real-time and more suitable for complex electromagnetic environment.In recent years,Convolutional Neural Network(CNN)has been widely used in various interdisciplinary subjects.With the characteristics of complete feature extraction and small detection error,CNN-based image denoising and object detection technology is suitable for more and more complex electromagnetic environment,which has become an inevitable means of SAR image processing research.Based on the characteristics that CNN can automatically learn and adapt to complex and changeable environments,this paper designs SAR image denoising and object detection algorithms based on deep learning.The main works are as follows:(1)A SAR image denoising algorithm based on deep learning and residual unit module is proposed.Aiming at the shortcomings of traditional filtering denoising algorithms,such as low efficiency and over-reliance on filter design,this paper designs an improved SAR image denoising algorithm based on residual network.The algorithm introduces a residual network module,which takes the noisy original image as input and the pure noise image as output.Through the mapping of input and output,the amount of parameter calculation and the difficulty of training are greatly reduced.When building the network training dataset,the Train400 dataset is first used to simulate the real SAR image by artificially adding multiplicative noise to approximate the image carrying speckle noise.In addition,the RSOD dataset is used to verify the denoising results of the algorithm on real SAR images to ensure the practical practicability of the algorithm.Experiments show that the algorithm achieves better performance than traditional algorithms and other deep learning-based algorithms in SAR image denoising tasks.It can effectively filter out speckle noise in a wide range and eliminate most of the noise spots in the image.It also makes the texture,edge and other details of the image not lost,ensuring the integrity of the image.(2)A deep learning object detection algorithm based on YOLOX is proposed.Aiming at the problem that the size of SAR image changes greatly and needs real-time detection,a SAR object detection algorithm based on YOLOX is designed.The algorithm introduces the adaptive activation function Meta-ACON,which improves the feature extraction ability of the Backbone.In addition,the Convolution Block Attention Mechanism(CBAM)is introduced to improve the feature representation and concentration ability of FPN for small size output,and end-to-end processing is realized.The NWPU VHR-10 dataset is used to verify the object detection performance.Experiments show that the algorithm has good detection performance for objects of various sizes in SAR images under the condition of complex electromagnetic wave scattering background.At the same time,it also meets the needs of real-time detection and realizes highprecision detection of SAR images.