Research On Terahertz Imaging Target Detection And 3d Reconstruction Algorithm

In recent years,with the increasingly complex international situation and the rising of new terrorism,it has become a development goal to find a substitute for X-ray for non ionizing radiation detection of people and objects.As an electromagnetic wave with a wavelength between microwave and infrared bands,terahertz has the characteristics of penetrability,low energy,and stable absorption spectrum of specific substances.It is widely used in non-destructive testing,human security inspection,biological tissue diagnosis,military detection and other fields.Although terahertz has good physical characteristics,due to the limitations of terahertz generation and acquisition device maturity at this stage,problems such as low imaging resolution,blurred image boundary and noise caused by diffraction fringes,overlapping textures before and after transmission images,and reduced image clarity will occur,which cause great difficulties for terahertz image recognition,splicing,and three-dimensional reconstruction.Therefore,based on the white X-ray migration data set and a small number of terahertz transmission experiment data sets,this paper proposes to use the difference of homography transformation between the front and back textures of the multi view transmission image to reconstruct the multi view cost volume,combined with the depth implicit model and multi-scale network,A transmission texture separation algorithm based on MDEQ(TTS-DEQ,Transmission Texture Separation by Deep Equilibrium Models)is proposed to partially preprocess terahertz transmission images.It provides a theoretical basis for THz image stitching and 3D reconstruction.At the same time,based on the analysis of terahertz imaging characteristics,an improved MVSNet 3D reconstruction algorithm based on Transformer(Feature and Cost Transformer Depth Inference for Unstructured Multi view Stereo,FCTMVSNet)is proposed.The self-attention mechanism is used to replace the traditional convolutional feature extraction network to solve the problem that the traditional 3D reconstruction algorithm feature extraction network is limited by spatial location information and insensitive to global information.At the same time,the inter layer attention mechanism of cost body is proposed to improve the network accuracy,which provides a theoretical basis for the application of algorithm engineering.Use structured position features to slice and output the 3D model obtained in the process,and realize specific object detection.Through experimental verification,the algorithm basically meets the design requirements.