Research On Weak Target Detection Based On High Sensitivity Infrared System

The infrared detection system with high radiation sensitivity is widely used to detect weak radiation targets,for example,steel defects detection based on surface emissivity differences.High radiation sensitivity is the prerequisite for infrared detection system to capture weak radiation target signals from complex backgrounds.Effective background suppression method and weak radiation target signal extraction algorithms are the key,which directly affect the final detection effect.The complexity and real-time performance of the algorithm determine its practicality on resource-limited hardware.Under the precondition of high radiation sensitivity,this thesis mainly studies infrared image preprocessing algorithms?including infrared image non-uniform correction method and infrared band image generation method?,complex background suppression algorithm of infrared image,infrared weak target detection algorithms?including texture filtering methods and intelligent methods based on neural network?,model compression methods used to improve the practicality of the algorithm,the specific research work is as follows:?1?The multi-frame accumulation method is used to suppress noise and improve the radiation sensitivity of the infrared detection system.Experimental verification has been done.The signal-to-noise ratio of target radiation relative to background radiation is used to quantitatively measure the weakness of the signal.?2?Aiming at the non-uniformity of infrared grayscale image caused by the difference of response rate and dark current between multi-element detectors,Dense-CNN correction method based on densely connected convolutional neural network is proposed,which breaks through the limitation of linear distortion theory modeling in traditional methods.In this dissertation,the peak signal-to-noise ratio?PSNR?of Dense-CNN method is respectively 149.64%and 88.24%higher than the combined wavelet-Fourier filter method?WD-FT?and median histogram equalization method?MHE?.Dense-CNN method solves the problem of blurring details caused by WD-FT and MHE correction methods.?3?Aiming at the problem of insufficient sample size of infrared band data,an infrared image generation method MSP-pix2pix?ssim-mae?based on generative adversarial network is proposed.A large amount of image data is needed in the study of detection algorithm.In order to expand the sample size of infrared image data,the infrared image is generated by using generative adversarial network technology.The network architecture and network loss function are further optimized.The experimental results show that the visual effect of the infrared image generated by the MSP-pix2pix?ssim-mae?method in this dissertation is better than the linear method,and the objective quality index PSNR of the infrared image is improved by 48.39%compared with the linear method.?4?In order to suppress the complex background of infrared image,an isotropic weighted soft threshold filtering method based on wavelet decomposition and reconstruction is proposed.The infrared detection system with high radiation sensitivity has the ability to capture the weak radiation target signals.At the same time,the interference from complex background is more prominent.In this dissertation,taking the specific infrared weak radiation target?lake surface vortex?as example,the texture feature value is extracted based on the gray level co-occurrence matrix of the image,the texture direction and law of the background and vortex area are analyzed,and isotropic weighted soft threshold filtering method is proposed.Compared with other background suppression methods based on frequency domain filtering and spatial domain filtering,the isotropic weighted soft threshold method in this dissertation can effectively suppress the background?the background roughness index is reduced by 77.32%compared with the average method?.After background suppression,the signal-to-noise ratio is improved by 9.79db.The problem that the frequency and spatial filtering methods cannot maintain the vertical texture of the vortex signal is solved.?5?Aiming at the detection of infrared weak radiation targets,a Gabor FM method based on texture filtering is proposed.The detection index F₁ score is 89.66%higher than that of image binarization method,which solves the problem that the image binarization method based on gray features can not avoid the interference of the pixel areas from background.In order to further improve the detection precision aimed at infrared weak radiation target,an intelligent detection method SMDU-Net based on saliency and multiple receptive field feature fusion network is proposed.Aiming at the characteristics of infrared image,such as low definition and low contrast,the design of neural network architecture is optimized to enhance the global information utilization and feature extraction ability of the network for infrared image.Aiming at the weak characteristics of the target,the feature map with the ability to highlight the target area is acquired by visual salience and input into the network,so as to improve the detection ability of the network for weak radiation target.Comparative experiments based on infrared vortex data sets show that the average precision?AP?of the SMDU-Net method in this dissertation is improved by 16.85%compared with the classical U-shaped network method,and the detection index F₁ score is 36.70%higher than that of the traditional texture filtering method.?6?Aiming at the problem of insufficient practicality for complex algorithms on resource-limited hardware,a model compression method Res MBS-Net based on mixed binary weights and residual connection is proposed.The weak characteristics of target increase the difficulty of detection and the complexity of the algorithm.Considering the portability of the algorithm on resource-limited hardware platforms?such as satellite-borne and airborne equipment?,this dissertation proposes Res MBS-Net method for the compression and acceleration of algorithm model.The Res MBS-Net method in this dissertation has the ability to greatly compress and accelerate the ordinary neural network model?the operating efficiency is increased by 6.29 times?,which improves the hardware portability of the model.Compared with the global binary compression and acceleration method?GBS-Net?,the average precision?AP?by the Res MBS-Net method is improved by 46.42%,effectively solving the problem of poor detection precision caused by the global binary method.