Infrared Small Target Detection Based On Mathematical Morphology And Robust Principal Component Analysis

As a crucial technique in infrared imaging field and automatic target recognition system, infrared small target detection is widely used in domains such as precision guidance, search and rescue acts, forest fire early warning etc. As a result of far operating distance, complicated backgrounds and heterogeneous noises, small targets occupy few pixels, lack texture, size, profile and other characteristics and difficult to detect. Previous studies meanly focus on areas such as filtering, neural network, wavelet, etc. The fact that the existing algorithms need a large amount of calculation makes them hard to employ in practical applications. Therefore, researches on real-time algorithms of detection for dim point target in low SNR and SRC IR pictures are of great scientific and practical importance. In this paper, several relevant concepts and application background are briefly introduced. It also reviews the key points, current problems, and overseas and domestic research status with regards to mathematical morphology and robust principal analysis. The main works and innovations are outlined as follows:1. Combined with the merits of morphology and surface fitting, a small target detection method based on NWTH transform and elliptic paraboloid fitting is proposed. Firstly we arrive at a conclusion that single structure element is deficient in infrared dim point target detection by referring concisely to grayscale morphology with emphasis on Top-hat operator. Then a modified Top-hat transformation named NWTH which can process background and clutter spontaneously by using two structure elements of different sizes and same shape is adopted to suppress background, its characteristics are analyzed specifically. After a study of infrared imaging system, target is considered as convex hull which can be approximated to an elliptic paraboloid. Finally, method of least squares is employed to further enhance real dim target in cluttered background. The presented algorithm has been testified to meet the real-time demand and with simple threshold segmentation, the position of small target can be predicted easily and effectively in different extreme cases such as complex background, strong noise and target invisible to naked eyes.2. The feasibility and validity of robust principal analysis theory for detection ofinfrared target are studied by introducing the processing and analysis methods of large and high dimensional data. As for an infrared picture, it can be treated as a superposition of three pictures(targets, background and noise). Theoretical analysis proves that clutters in the IR images meet 2 conditions: 1.independent and identically distributed Gauss distribution; 2.sparsity and large amplitude. Computer simulations testify that the background images are low rank matrixes. Hence, we can come to a conclusion that RPCA can be applied in infrared small target detection. On the basis of the fact, a method which adopts APG, DUAL and EALM respectively is designed to verify the effectiveness. The method first estimates background, then original image subtracts predicted background to obtain residual image, finally it selects a proper threshold(the product of grayscale average and a ratio) to detect target. By seriously analyzing and comparing simulation results from 4 completely different aspects(gain of SNR, detection time, threshold coefficient selection and subjective visual quality), a conclusion is drawn that APG shows better performances than DUAL and EALM in infrared dim target detection field.