Underwater Target Recognition Method Based On Polarization Imaging And Deep Learning

Underwater target recognition technology is an important guarantee for underwater works,which plays an important role in both military and civil fields.According to the different information acquisition methods,underwater target recognition technology is divided into underwater acoustic detection technology and optical vision detection technology.Compared with underwater acoustic detection,optical vision detection has many advantages,such as intuitive target detection,high resolution,excellent real-time performance and so on.However,water and impurities have strong absorption and scattering effects on optical signals,which seriously damage the clarity of underwater optical vision detection and limit the underwater target recognition distance.Therefore,how to effectively reduce the impact of water and impurities and improve the clarity of underwater imaging has become a difficult problem and development direction in the field of underwater target recognition.In order to solve this problem,this paper uses the method based on polarization imaging,combined with image fusion and image processing technology based on deep learning,respectively improve the frontend hardware and back-end processing of underwater imaging system,and finally achieve clear underwater imaging and accurate target recognition.The specific work contents are as follows:In order to solve the problems of poor image definition and limited underwater detection distance caused by strong absorption and scattering of light signal by water and impurities,this paper analyzes the defects of traditional polarization imaging in underwater imaging denoising based on the mechanism of absorption and scattering.Then,on the basis of theoretical analysis,the scattering and polarization characteristics of underwater imaging are simulated in ZEMAX,and the modulation law of polarization direction caused by the change of water turbidity and underwater target surface roughness is obtained.Finally,according to the law of simulation,the multi polarization imaging combined with image processing method based on deep learning is proposed,and the principle model of underwater multi polarization imaging system is established.For the image processing part,firstly,the multi polarization images are fused,and then the method based on deep learning is used for further processing.Among them,image fusion is based on wavelet transform.In this paper,an underwater multi polarization image fusion algorithm based on wavelet transform is proposed to suppress the influence of backscattering and improve the image contrast.On this basis,based on the principle of deep learning image processing,a Generative Adversarial Networks(GAN)based is built and trained based on Tensorflow.After training,the underwater image can be processed,which can reduce the influence of absorption and forward scattering,and get a clearer underwater image.Based on the above two works,an underwater multi polarization imaging system is designed and developed.In the hardware of the system,the technology of fly-eye lens is used to realize uniform linear polarized lighting,and FPGA is used to control the stepping motor to switch the polarization state and trigger the camera to synchronously capture images.Then,based on Visual Studio,the control software is designed with C#,which include lighting control,polarization state switching,image preview and acquisition functions.Finally,the underwater imaging verification experiment is carried out.The experimental results show,the peak signalto-noise ratio(PSNR)is about 6% higher than that of conventional underwater imaging,and the observable water turbidity(NTU)can be increased by 20 at the same distance and imaging clarity.