Research On Stripe Noise Removal Algorithm For Infrared Images

In today's highly developed image-forming system,infrared images play an important role in it.Since the infrared image-forming system can sense the infrared thermal radiation of the target by the infrared detector and invert the target temperature for imaging.Exploiting the difference in temperature of each object,it can greatly compensate for the shortages of visible image-forming system at night and in special scenes.It plays an important role in military reconnaissance,scene monitoring,autonomous driving,night vision systems,fire detection and other fields.The infrared image stripe noise removal is one of the important research topics in the field of infrared image processing,which has important practical significance for the subsequent research in the fields of infrared image recognition and classification.Infrared imaging systems have a large number of non-uniformities due to the level of technology and time of development.These non-uniformities further lead to the generation of stripe noise.In recent years,many scholars at home and abroad have devoted themselves to the study of non-uniformity stripe noise removal from infrared images.However,due to the special structure and direction of the infrared stripe noise,it is difficult to remove the stripe noise directly,and in the process of removal,it is easy to cause the blurring of the edge information parallel to the direction of the stripe or the loss of detail information in the non-stripe noise region.These challenges make most current infrared image stripe noise removal methods inadequate.In this paper,we will start from the a priori knowledge of stripe noise,and fully combine the ideal image's own characteristics,to build an infrared image stripe noise removal model and network,and carry out research work on infrared image stripe noise removal methods.The main research contents and innovative results are as follows:(1)Currently,most infrared image stripe noise removal algorithms do not sufficiently exploit the a priori knowledge of stripe noise,and it is difficult to effectively distinguish between stripe noise and the edge features parallel to it,which cannot balance the removal of stripe noise and the protection of image edge information.To address this problem,a stripe noise removal algorithm based on adaptive edgepreserving operator(AEPO)with multi-sparse representation is proposed.The advantages of this method lie in digging into the global sparsity of stripe noise and the sparsity of gradients in the stripe direction,as well as the sparsity of gradients in the vertical stripe direction of the ideal information image,constructing a stripe noise removal model,and then adaptively adjusting the edge retention operator according to the difference in pixel edge contrast to improve the protection of edge information parallel to the stripe direction.It can achieve reasonable removal of stripe noise in infrared images and effective protection of edge information.The experimental results show that this algorithm not only achieves excellent denoising effect(the NR index is improved by 8.23% on average in the experimental images),but also realizes effective protection of edge information(the MRD index is improved by 2.01% and the ID index reaches 0.992 on average in the experimental images),and obtains ideal denoising performance.(2)Currently,most model-based stripe noise removal algorithms for IR images tend to use global modeling of the entire image and optimize it,but this approach inevitably results in pixels in non-stripe noise regions being affected to some extent,and even blurred details or image distortion.To address this problem,a multi-scale wavelet transform with multi-sparse representation(MWMS)based stripe noise removal algorithm for infrared images is proposed.The advantage of this algorithm is the ability to concentrate the stripe noise in the vertical direction by multi-stage multiscale wavelet transform with different intensities in different levels of vertical components,and protect the other information from the model in other components.Moreover,different regular term models are chosen for the stripe noise of different intensities in the vertical components of different levels to ensure that the stripe noise in the vertical components of each level is maximally rejected and eventually reconstructed by wavelet inverse transform to obtain the ideal denoised image.The experimental results show that this method has good denoising ability(the NR index is improved by 7.69 % on average in the experimental images)and excellent protection of information in the non-striped noise region(the MRD index is improved by 4.23%and the ID index reaches 0.994 on average in the experimental images),and satisfactory denoising results are obtained.(3)To address the shortcomings of most current neural network-based multi-frame continuous image stripe noise processing algorithms,which cannot balance the stripe noise removal of moving pixels with the protection of stationary pixels from excessive smoothing and are prone to "artifacts",we propose a multi-frame IR image stripe noise removal algorithm based on adaptive motionstates and neural networks(AMSNN).The advantage of this algorithm is that it can apply different update strategies and adaptive iteration rates to the correction parameters and bias parameters for removing stripe noise for motion pixels and still pixels for the different needs of denoising in that frame,ensuring that the still pixels will not be over-smoothed and the motion pixels will get a better denoising effect.The experimental results show that this algorithm has strong denoising ability(the NR index is improved by 5.05% on average in the experimental images)has good information protection of still pixels(the MRD index reaches 3.448 and the ID index value reaches 0.9915 on average in the experimental images),and obtains excellent denoising results.