Pixel-level Image Enhancement And Registration Algorithm

Image registration is an important step in multi-source data fusion, timing imageanalysis, target detection, pattern recognition, image mosaic and so on. With the imageacquisition means and technology continues to expand and improve, the challenges of lowaccuracy, low success rate and poor real-time faced by image registration technique is alsoincreasing. The research work is expained according to the logic mainline of pretreatmentprior to registration, homologous image registration and heterologous image registration.For the purpose of building a high-precision, high success rate and fast registrationalgorithm, in a large number of infrared, visible and multispectral image data support, thekey technologies encountered in image registration are analysed and researched. At thesame time, the pixel/sub-pixel registration algorithm overall design are discussed.Multi-level, multi-angle analysis of the experimental results are used to validate thealgorithms. The main work and innovations can be condensed as follows:To improve the image registration accuracy and success rate, three kinds of imagepre-processing programs are researched. First, against the weak sharpness and contrast ofthe image registration, an adaptive image enhancement algorithm of gray and details is putforward. The algorithm combine human visual system with the classic image enhancementalgorithms, use statistical characteristics of gray and gradient to enhance image adaptively.Secondly, against the problem of blurred edge and texture, a effective enhancementalgorithm based on edge saliency are designed. The algorithm is designed to improvecontrast of edges and textured areas and for the subsequent use of heterologous imageregistration by edge feature. Finally, in order to remove random noise and impulse noise, amedian filtering algorithm combination of mean filtering is put forward.The algorithmchanges the size, shape and filtering method adaptively according to the scenecharacteristics and the distribution of image noise, and can remove the redundantinformation in amplitude spectrum in visual attention model.Multi-sensor images registration based on characteristics. First, against the poorrobustness or high complexity of the existing descriptor, a novel descriptor is proposedbased on the gradient vector. This descriptor inherites the advantages of classic featuredector and overcome its scaling sensitive shortcoming, it is low-dimensional and inaccordance with human vision bionics. In addition, a new registration algorithm is designed based on above descriptor. It is very robust for translation, rotation, and has a certain scaleinvariance. For large-scale geometric distortion, registration still achieve100%success rate,and even achieve sub-pixel accuracy. Meanwhile, for the poor robustness and real-timeproblem of large area image registration, a new fast image registration algorithm based onthe sub-image features is proposed. With the relatively small distoration, the sub-imagewhich has strong contrast, clear structure extracted are used for image registration insteadof the whole image, that improved the accuracy and speed of registration.Multi-sensor images registration based on area. First, from a summary of theimage scene features analysis, the traditional registration method FMT limitations in highpass filter to extract the high frequency information is pointed out and a novel E-FMTregistration algorithm based on binary edge image is proposed. The E-FMT registrationalgorithm showes the robustness and real-time in images which have distinct edges.Secondly, against the existing problems in registration algorithm based on gray, amulti-sensor images registration algorithm based on cross-correlation of the edge region isproposed. The new algorithms is improved from three aspects of associating featureextraction, similarity function selecting and optimize algorithm, that has made registrationaccuracy reaching sub-pixel and improved the registration speed.Finally, for a class of visually significant image, a novel heterologous imageregistration approach based on visual attention is presented. Firstly, a visual attentionmodels based on spectral separation is proposed, which is segmented and extracted by themethods of morphology. The model is more accurate and complete than the classical visualmodel in the extraction of salient region. Meanwhile, the concept of minimum correlationimage is presented, instead of the whole image for alignment, that improved registrationspeed.