Few-shot Knowledge Transfer Learning Quantitative Technology And Experimental Verification Research Of Infrared Nondestructive Testing

In the fields of aerospace,rail transit,military,and pipeline transportation,many materials have near-surface and sub-surface defects,and non-destructive detection technology is required to repair.Carbon fiber reinforced plastic(CFRP)have the advantages of high intensity and low density.Long-term use will produce defects such as delamination and stratification to hidden safety hazards.The coating material has defects such as cracks and adhesive failure,which can affect stealth performance.The empty cave and other defects of glass fiber materials will cause the structural stiffness and intensity to decrease,and the protection performance and thermal insulation performance will not reach the standard.These problems may cause safety accidents.Optically pulsed infrared nondestructive testing is suitable for the detection of near and subsurface defect detection of most materials,and has the advantages of no contact,and fast and intuitive imaging.However,in practical applications,there is a signal processing problem: defect size quantification is a challenge.The existing algorithm can only semantic segment a single type of test piece,and the differences in different types of test parts are large;the data collection is difficult and the data volume is serious,reflex and other interference.The difference between defective areas and non-defective areas,infrared image defect information is covered with high noise and material reflection.In response to the above problems,the Thesis proposes an end-to-end few-shot learning semantic segmentation network,which can solve the problem of insufficient data volume and generalized performance.By using the prospect background guidance,it provides guidance information for query branches,and increases the measurement distance of defects and non-defect characteristics in order to better detect the defects.The research content of the thesis is as follows:(1)The thesis proposes the semantic segment algorithm of the domain adaptive fewshot learning defect,reduces the difference between the category between the support set query set,reduces the difficulty of similar information between the network learning support set and the query set,increasing the convergence speed of the few-shot learning network.Improving the ability to excavate difficult samples,the support image of the domain alignment can better provide guidance information.(2)The thesis proposes the query branch structure of the joint guidance of the foreground background,which can better provide guidance information.The memory guidance module is introduced.The memory guidance module can record multiple memory prototypes,and can provide more detailed information for query branches.Through the update of the memory module,the background guidance information provided is not limited to the current support image.The memory module can remember the multiple backgrounds of the historical support image.Increase the measurement distance between the prospects and the background through joint guidance by the foreground background,and reduce the missed errors of the detection.In response to the memory module,inter-class loss and intra-class loss are proposed,so that limited memory modules can store more information and ensure the diversity and accuracy of memory prototypes.Realize the memory prototype with small distance within classes and large distance between classes.From the visualization of defect detection,the objective evaluation of the evaluation indicators of the IOU and F-score evaluation indicators,the comparison of ablation experiments,and the same method comparison.Analyze the effectiveness and robustness of this algorithm.Finally,from the perspective of actual application,this algorithm is deployed on portable infrared detection equipment to achieve automatic segmentation and quantification.