Infrared Weak Small Targets Detection Based On Multi-scale Geometric Analysis Method Study

Infrared dim target detection plays an important role in remote surveillance,infrared warning, infrared target recognition and tracking system, and so on, and anin-depth discussion and research of their methods has cruial meaning and practicalvalues. In this thesis, based on the Multiscale Geometry Analysis (MAG) theory, somenew infrared dim and small target detection algorithms are studied and provided, themain work as follows:(1) The history of development about the adaptive and non-adaptive MAG isresearched, and basic principles and characteristics of various tools are given. As wellas, the commom algorithms of infrared background suppression are analyzed byexperimental data, including the Max Median Filter, the Local Gray LevelCharacteristic, and the Gray Level Morphology.(2) An algorithm based on the nonsubsampled contourlet transform (NSCT) ispresented for target detection. First, the NSCT is used to decompose the infrared imageinto various scales and directions, and these NSCT coefficients at varioushigh-frequency subbands are modified to eliminate the influence of noises andbackgrongd edges, and the low-frequency subband is maked zeros to eliminatebackground. Then the modified subbands are resconstructed, the target detection isrealized using adaptive thresholding segmentation finally, and it is better than the targetdetection algorithm based on the classical wavelet transform.(3) A method based on multiscale features fusion in the shearlet transform domainis proposed for fast detecting the dim target in infrared images. First, an original imageis decomposed using the shearlet transform with translation invariance. Second,according to the different characteristics of modulus maximums of the shearletcoefficients at various scales and directions, the background and noises of infraredimage are suppressed and targets can be enhanced by fusing the multiscale andmultidirection shearlet coefficients. Finally, the dim targets are detected bythresholding segmentation with a maximum contrast criterion. The experimentalresults show that the proposed method is of good performances.