Visual Saliency Computation With Its Applications In Infrared And Visible Image Fusion

The development of information fusion technology makes it possible for many sensors to work together,Infrared and visible sensor fusion is one of the most important branches of it.Image fusion integrates the images from different sensors in order to obtain more abundant and useful information,and improves the integrity and accuracy of the scene description.Among them,“which part of the image needs to be fused,what to adopt the fusion strategy” is the key to the fusion algorithm.The introduction of visual saliency mechanism enables the system to focus only on salient regions of the image,while reducing or ignoring non salient regions,according to different regions,with different fusion strategies,which not only can reduce the amount of information computation,and improve the effect of image fusion.Aiming at the blur and the low contrast characteristics of infrared images,a method combining frequency domain saliency and spatial saliency is proposed to detect infrared salient targets.The method uses both local and global information of the image,so that the significant target information in the spatial and frequency domain can be complementary,so as to obtain more accurate infrared saliency map.Among them,the shape feature of infrared target is extracted by the combination of edge detection operator and mathematical morphology operator,it enhances the outstanding characteristics of infrared target and suppresses the effect on infrared background in the spatial saliency map;the frequency and spatial saliency map are fused into a global saliency map which guides infrared salient target detection.This method provides a new visual saliency computation for infrared image,and can be used to calculate the infrared saliency map in the following chapters.Visual saliency is helpful to quickly locate important information in the scene.An infrared and visible image fusion method based on saliency controllable weight strategy is proposed though applying visual saliency to the field of image fusion.First of all,considering the imaging characteristics of visible image,the improved LBP operator is used to calculate the saliency of visible image.Then,though the competition of the saliency between infrared and visible images,avoiding the problem that considering only the saliency of infrared image,expanding the applicable range of the fusion algorithm.Through the regional average saliency of the salient regions is calculated,and the anti-noise performance of this algorithm is improved.Finally,through the saliency measure of the salient regions guiding the fusion process,a new region feature description and a weighted fusion rule are proposed.This method has fast operation and good real-time without prior knowledge.The scale of the target and the characteristics of target in multi-target image can be not known in advance.For this problem,the best scale of different objects is selected by the introduction of the concept of saliency scale-space,and combining the three aspects of scale,saliency and fusion rule,a fusion algorithm of infrared and visible images based on saliency scale-space in the frequency domain is proposed.On the one hand,the spectrum scale space is established by Hypercomplex Fourier Transform(HFT)and multi-scale Gauss kernel function,and the salience of different scale targets is detected by selecting the best description scale according to the principle of minimum two-dimensional entropy;On the other hand,aiming at the difference of multi-target salient regions in infrared and visible image,the salient region correlation map is generated by correlation processing,and guides fusion processes of different regions.This method is applicable to complex scenes with multi scale significant targets.In view of the possible block effect in spatial fusion methods,an image fusion method based on region saliency and the translation invariance,multi-scale and multidirectional characteristics of NSCT is proposed.According to the regional association mapping,the target regions and background regions are divided,and their different fusion rules are developed based on region features.Then the low frequency and high frequency subbands coefficients of NSCT transform of different regions respectively are fused,and the fusion image is obtained by inverse NSCT transform.This method makes full use of region features,different regions and different frequency subbands adopt different fusion rules.Compared with the traditional methods,this method is more in line with the actual requirements and has better visual fusion effect.