Research On Forward-Looking Sonar Image Sequence Mosaic

In underwater detection applications,two-dimensional imaging forward looking sonar(FLS)is the main tool to obtain the visual information of scenes.However,due to the limited field of view,the FLS images cannot present the overall view.It is not convenient for visualisation or computer techniques to deal with a partial view of an object in a frame.It is therefore important to integrate multiple frames,with overlapping regions,into one composite image that offers an overall view of the filed and high resolution of interest without any seam.In this dissertation,the FLS image sequences are used as the research object.The frames with a limited field of view are integrated into a panoramic image,which adequately reflects the underwater visual information and provides a technical means for underwater image mosaicing.This dissertation focuses on the main steps of image mosaic,including image registration,global alignment and image fusion,and then organizes them into a completed FLS image mosaic algorithm.The main work and achievements of this dissertation includes:1.An FLS image registration algorithm based on local region phase correlation for near-consecutive frames is proposed.The transform between the FLS frame satisfies the rigid transformation,including a rotation and a translation vector.Commonly,there is no significant feature in the FLS images,so the feature based registration methods cannot satisfy the requirement of the registration of the whole image sequence.All the pixels in some image regions are better to be used for image registration.However,some characterizations of the FLS images,such as poor gray values,low SNR and the uneven distribution of illumination,all influence the effect of the registration.To solve these problems,the proposed registration algorithm utilizes the original data of FLS images to estimate the rotation angle,and does not involved the interpolation for polar transform,thus overcoming the influence of the poor quality of the FLS images.And the phase correlation method is robust to noise and can be used to calculate the registration parameters of FLS images.In addition,the proposed algorithm can suppress the influence of uneven illumination by selecting the registration area adaptively.In the experiment,the accuracy and real-time has been analyzed using the FLS image sequences of different scenes,which shows the superiority of the proposed algorithm.The proposed registration algorithm meets the needs for FLS images mosaic.2.A global alignment method based on pose graph optimization is proposed.Firstly,the definition of the variables in pose graph is specified and the pose graph of FLS image sequence is established.To eliminate the accumulated registration errors,the images in theloop closure are expected to be aligned to establish the global constraints.However,the uneven illumination has an obvious influence on the registration in the loop closure condition,and the near-consecutive registration algorithm cannot satisfy the image registration under this condition.To solve this problem,the linear features in the frames are extracted for image registration in loop closure.The proposed method enhances the directional information of the image to highlight the linear feature of the FLS images.Then the Hough transform is used to extract the line features of the image,and the selected lines are used to register the frames in the loop closure.The registration results are added into the pose graph as global constraints,and the global alignment of the mosaic can be ensured by the graph optimization.3.An image pixel-level fusion algorithm is proposed.The algorithm can be divided into coarse and fine steps.Due to the particularity of FLS imaging,the definition of the sharpness is different from the sharpness description of the optical images.Aiming at this problem,Gabor energy and local orientation contrast are proposed as the fusion rules.In the coarse fusion step,the source images are first decomposed into multi-scale sub-bands using the non-subsampled contourlet transform(NSCT),then the sub-bands of the source images are merged,based on the Gabor energy and the local contrast for low and high frequency sub-bands,respectively,to generate the fused sub-bands which are used to produce the fused image by inverse NSCT.In the fine fusion step,a decision map is used to choose the pixels of the fine fused image from the coarsely fused image or one of the source images.This decision map is first constructed by measuring the similarity of the coarsely fused image to the source images,and then processed by a morphological post-processing technique to ensure its continuity and smoothness.Extensive experiments on FLS image fusion and mosaic have been conducted to demonstrate the effectiveness and the superiority of the proposed technique using both a subjective evaluation and objective metrics.4.The FLS image mosaic algorithm is studied.The algorithm includes FLS image registration,pose graph optimization and FLS image fusion,and combines the track information recorded by the FLS.The algorithm is initially aimed at the simplest motion condition of the FLS image sequence.With the increasing complexity of the motion,the algorithm is improved and generalized to more motion form of FLS.The experimets includes only rotation,rigid transform and loop closure FLS image mosaic.The results demonstrate that the proposed algorithm can be adapted to the FLS image sequence mosaic in a wide range of motion forms.This dissertation investigates the FLS image sequence mosaic algorithm,including FLS image registration,pose graph optimization and FLS image fusion.The proposed techniquesare tested using measured data and the expeirmental results show that the proposed method can promise good accuracy and stability,which illustrates that the algorithm is suitable to the FLS image sequence and has high engineering value.