Small Dim Moving Target Detection Based On Spatial-temporal Sparse Representation

The technique of infrared small dim moving target detection is a key technique ofoptical imaging detection system search. The target in the deep space is far away fromthe observing system, and the target is usually small and weak in the imaging detectionsystem, and moreover, it is submerged in various complex noise and clutters, whichmake target detection difficult. Therefore, the detection method of infrared dim targetwith low SNR is very important.Recently, sparse representation theory attracts huge attentions, it uses fewer atomsof the over-complete dictionary to reveal the main content of signal, which makes signalrepresentation more accurately and effectively. Although the sparse representationtheory is still not fully perfect and needs for further improvement, it has achieved manyresults in the field of signal processing, and shown a great potential.Based on the above background, the small dim moving target detection method inthe image sequence is deeply researched using sparse representation theory in the thesis.Firstly, time domain characteristics of the image sequence are analyzed, theover-complete dictionary from single frame and sparse representation can’t excavatemotion information of target in image sequence. Then, base on physical properties ofcontinuity and consistency of motion of the target, this paper combines with spatial andtemporal domain to train spatial-temporal over-complete sparse dictionary, explores thecharacteristic differences of target and background, and then puts forward the algorithmof small target detection based on spatial-temporal sparse representation. In thealgorithm, the spatial-temporal over-complete dictionary is constructed by regarding theimage sequence as training samples, the decomposition coefficients of image blocks areobtained in the spatial-temporal sparse dictionary, and then the image blocks aredetermined which the image blocks contain target or not by combining the coefficientsand the content of atom corresponding to the large value. In order to further improvetarget detection performance, the spatial-temporal over-complete sparse dictionary isclassified as the target spatial-temporal and background spatial-temporal dictionaries byusing Gaussian dictionary, and decomposition coefficients of the image blocks with thedictionaries are obtained. The target detection is completed by residual difference thatare separately from the reconstructed of the target and background spatial-temporaldictionaries. The experimental results show that spatial-temporal sparse dictionary overcomes the limitation that spatial sparse dictionary can only describe themorphological characteristic of target; the spatial-temporal classification dictionary canfurther enhance the sparse feature differences of target and background signal, which ismore helpful to detect target.