Research On Algorithm Of Registration And Fusion Of Infrared And Visible Light Images Of Blower

Infrared and visible light image are widely used in the field of industrial blower.Infrared image can be used to detect temperature anomalies,and visible light image can be used to observe external characteristics of faults.With the fusion of those two types of image,engineers are able to determine the fault location and type more accurately,which helps them to detect,diagnose and monitor operating conditions of blower much better.Since traditional image registration algorithm has low accuracy in registering infrared and visible light image of blowers,the two types of image can not fuse together efficiently.Therefore,this thesis proposes three algorithms to achieve good fusion of them.The main contribution and innovation are in the following :(1)To improve the registration accuracy of traditional algorithms for infrared and visible light image of blowers,this thesis proposes W-GSCD(Weight Global Shape Context Descriptor)and KSC(K-Means++ and Slope Consistency)blower image registration algorithm based on contour segmentation.Firstly,prepropcess the original images to obtain the datasets of blowers for the algorithms.Then,use an improved Deep Labv3+ segmentation network to extract the target contour with high precision by embedding the Mobile Vi T module and adding the Boundary Loss.Next,employ the Harris corner detection algorithm to extract contour feature points as objects for subsequent feature description.The W-GSCD algorithm is proposed to describe the extracted feature points,and the initial matching is completed by using a bidirectional registration strategy.Then,introduce the K-Means++ and Slope Consistency algorithms to screen feature point pairs,obtaining stable and high-quality matches,register the infrared and visible light image of blowers with high quality.(2)To enhance the fusion quality of registered infrared and visible light image of blowers,for example there are unclear salient regions and poor detail textures,this thesis proposes an end-to-end image fusion algorithm based on an attention mechanism.Firstly,preprocess infrared and visible light image of blowers to obtain the registered image.Then,improve the residual fusion structure through embedding Efficient Channel Attention Module(ECA)to enhance the region of interest and improve texture details.Finally,verify the effect of the ECA module and different loss parameters on the fusion of images with ablation experiments.(3)For it requires manual preprocessing to solve the problems of the poor quality and ghosting caused by unregistered infrared and visible light image of blowers,this thesis proposes a three-stage network-based image registered fusion algorithm.Firstly,generate realistic pseudo-infrared image with a pseudo-infrared image generation network to transform the cross-modal image registration problem into a homologous one and reduce the difficulty of image registration.Then,register the pseudo-infrared image and the real infrared image via infrared image registration network to generate the registered infrared image,replacing traditional registration algorithms and reducing the cost of manual preprocessing.Finally,fuse the registered infrared image and the real visible light image with an end-to-end image fusion network of blowers.The result shows that the quality of the registered image is improved and the problem of ghosting is reduced.