Research On Lightweight Single Infrared Image Super-Resolution Reconstruction Method

Infrared thermal imaging has the advantages of long-distance imaging,strong penetration,good anti-interference,working at all-weather and all weather,non-contact,good concealment,high sensitivity to object thermal radiation and so on.Thus,it can be widely applied in military and civil fields.However,due to the weak thermal radiation of the imaging target itself and the limitations of the infrared detecto r,the resolution of the infrared images output by the existing imaging systems is usually lower than that of the visible image,and the quality of the infrared image is often low.There are some shortcomings,such as low signal-to-noise ratio,low contrast,blurred edge,lack of detailed information.Therefore,how to effectively improve the resolution and quality of infrared images has inevitably become a problem for researchers to address urgently.In recent years,the quality and clarity of low-resolution images can be greatly improved via deep learning-based image super-resolution technology without breaking through the complex hardware manufacturing techniques and high cost.Therefore,to address the problems existing in the single infrared image super-resolution task,this thesis takes the low-resolution infrared image as the research object and focuses on the construction of lightweight model based on the deep learning theory.The main contents and contributions of this thesis are listed as follows:(1)Considering the problems of low contrast and lack of detailed information of infrared images,as well as how to extract features efficiently,this thesis improves the repeated up-and down-sampling mechanism to extract and refine features.Combined with the attention mechanism,a novel self-corrected attention block(SACB)is proposed.Taking SCAB as the basic construction module,a lightweight infrared image superresolution algorithm based on self-corrected attention mechanism is presented.In addition,an infrared imaging system is built to obtain enough infrared images for experimental verification.The system is lightweight and portable,which is mainly built by a personal computer and infrared module.Sufficient quantitative and qualitative experimental results show that the performance of the proposed method is superior to the comparison model,which not only greatly reduces the amount of parameters and computational cost of the model,but also can run quickly.(2)Considering how to achieve efficient multiple feature extraction and expression,as well as the weight distribution and fusion of different features,this thesis further improves the algorithm and proposes a novel multi-path feature fusion block(MFFB).For the first time,MFFB can explore the effects of linear features,non-linear shared-source features,and attention features on the performance of infrared image SR.The effects of nonlinear residual features and attention features on the performance of infrared image super-resolution reconstruction.The ablation experiment results demonstrate that each feature extraction branch of MFFB contributes to improving the feature representation ability of the model.(3)Considering how to build the lightweight image super-resolution model,this thesis proposes a lightweight infrared image super-resolution network based on MFFB.The proposed algorithm not only utilizes the recursive mechanism to realize parameter sharing,but also introduces the feature feedback mechanism to refine semantic features and improve the representation ability of features.Many comparative experiments show that compared with the existing lightweight models,the proposed algorithm is effective in the image super-resolution task of multiple scale factors,and it can achieve an excellent trade-off between model performance and running time.In conclusion,the proposed lightweight infrared image super-resolution methods are tested and compared with the existing lightweight super-resolution algorithm on the test datasets from different scenes obtained by the infrared imaging experimental system.The experimental results demonstrated that the proposed methods can not only achieve superior performance on several no-reference metrics,such as Entropy,Laplace,Roberts,EOG function,SMD and SMD2,but also greatly reduce the number of parameters and calculation of the model.Furthermore,the proposed methods can effectively and quickly restore the details of infrared images,and effectively overcome the problems in the existing infrared image super-resolution methods.