An Adaptive Algorithm For Infrared Small Target Detection Based On Image Sparse Representation

The detection of infrared small targets is a hotspot and a difficulty in the domain ofinfrared photo processing. Because of the small dimension of the target, we cannotidentify the target by its geometry and texture information, besides, the weaken contrastbetween the target and the background makes the work of accurate detect much moredifficult. The sparse representation based on over-complete theory is raised in recent years,using the over-complete dictionary to represent the IR images, to achieve the sparse resultby sacrificing the complexity of the basis. However, it is difficult to solve the sparse resultfor the algorithms proposed before. The aim of this paper is to present a new IR smalltarget detection algorithm based on sparse representation but easy to solve the sparseresult.The construction of over-complete dictionary is one of the keys for the sparserepresentation theory. A proper dictionary should submit to:1. it could disassemble andreconstruct the IR image efficiently;2. the dictionary could not be too complex, or else thecomplexity of calculation will be burdensome. The construction of over-completedictionary could put into two different categories: unit two or more orthogonal basisexisting before, or construct the dictionary straightly according to the samples’ characters.Considering the second method is easier, more flexible, and closer to the characters of thesamples, this paper will adopt this method to construct the over-complete dictionary.Based upon the previous work, this paper will use the over-complete dictionaryconstructing a standard reference model, then pursuit the sparse result for the difference ofthe image representation and the reference model so the location and characters of thesmall target could be confirmed, to avoid the burdensome sparse result calculation.Improvements for the algorithm are also proposed in this paper considering thebalance between the accuracy and the speed, mainly include:1. Adopt TDI to improve theSNR of IR image and then improve the accuracy of detection;2. Screening doubtful targetzones according to the IR image’s energy distribution, to reduce the quantity of calculatework and improve the algorithm’s speed.Experimental results show that the algorithm proposed by this paper could detect thesmall target with satisfying accuracy when SNR is enough, as the SNR decreasing, theaccuracy of the algorithm decrease, too, but it still could identify the main location of thesmall target. Besides, the algorithm’s speed is fast especially when the IR image is large.