Atmospheric Scattering Model With Applications To Foggy Image Enhancement And Small Sea-surface Object Detection

In foggy scenes, imaging mechanism analysis, degraded image restoration and per-ceptionalgorithmresearcharequiteimportantforoutdoorcomputervisionsystems. Fundedby the National Defense973Project”Basic Research on Surface***”, this dissertation isfocused on two problems of foggy images:”how to remove” and”how to use”, where theformer is the imaging mechanism analysis and the enhancement algorithm research forfoggy images, while the latter is about the extended applications of fog effects and fog re-movalalgorithmsinsea-surfaceenvironmentperceptionsystems. Themaincontributionsand innovations of this dissertation are as follows.(1) The multi-scattering atmospheric model and the visibility restoration algorithmsof foggy images are studied in the dissertation. Firstly, to overcome the drawbacks ofsingle scattering atmospheric model, this dissertation studies the modeling method ofmulti-scattering atmospheric model, and provides the differences and relations of thesetwo models. Then, foggy image enhancement algorithms, which are based on the singlescattering atmospheric model and the multi-scattering atmospheric model respectively,are proposed. The former is mainly for structured scenes, such as sea-surface and roadscenes, while the latter mainly works for dense foggy scenes or scenes with highly dy-namicdepthranges. Experimentsshowthat,comparedwithmanywell-knownalgorithmsin recent years, the algorithms described above can make a better performance on visibil-ity restoration for road or sea-surface images.(2) A fast algorithm of small sea-surface visible object detection, based on atmo-spheric scattering model, is proposed in this dissertation. It is the production of a skillfulusage of the fog effects, mainly having two parts: the first one is the analysis of the darkchannel image (DCI) for sea-surface visible objects, and the rationality of DCI as a fea-ture extraction method for sea-surface visible objects; the second one is the study of howto use atmospheric scattering model to obtain depth-adaptive threshold estimation for ob-ject detection. Compared with the algorithms based on background modeling or saliencyanalysis, thealgorithmsdescribedabovehavemadeabetterperformanceaccordingtoourexperiments on sea-surface surveillance videos.(3)Thisdissertationproposesanovelsmallsea-surfaceinfrared(IR)objectdetectionalgorithm, based on the approximated wavelet transforms and the integration of multiple spatio-temporal vision cues. Firstly, in the viewpoint of”data microscope”, the disserta-tion provides the rationality analysis of DCI in small sea-surface visible object detection,and proposes a DCI induced background estimation method for sea-surface IR images.Then, based on the above work, the dissertation describes the idea of the approximatedwavelet transforms, and gives its application to the threshold estimation for IR objectautomatically segmentation. Finally, to obtain robust algorithms for small sea-surfaceIR object detection, this dissertation also provides a Bayesian inference based model formultiplespatio-temporalvisioncuesintegration. Experimentsonsea-surfacesurveillancevideos show that above algorithms can make more satisfactory results than many state-of-the-art techniques.(4) The dissertation has analyzed DCI in the viewpoint of wavelet transforms andstatistics respectively, and proposes two extended applications of DCI: the first one is theconstruction of data-driven wavelet filters for foggy image enhancement algorithms, andthe second one is the construction of order-statistic filters for background estimation ofsea-surfacevisibleimages. Theadvantagesoftheformeroneareitsabilitytomakesimilarresultsasthebilateralfilterswithfewercomputations,andthemoreskillfuledge-stoppingfunctions with O(N) computations. The advantages of the latter one are the real-time im-plementations of background estimation for highly dynamic sea-surface images, withoutdata association problems across frames. Whereafter, to obtain robust detection algo-rithms of small sea-surface visible objects, the dissertation analyzes the shared features ofobjects under various weather/lighting/backgound conditions, and proposes a Bayesianinference based statistic model, leading to a framework for clutter-adaptive sea-surfaceobject detection algorithms.Although the main algorithms in this dissertation are proposed for the sea-surfaceenvironment perception system, they can be extended to scene perception systems of au-tonomous vehicles.