Research On Infrared Small Target Detection Algorithms

Small target detection and tracking technology is of significantly importance in the defense and civilian areas. Compared to radar and laser, infrared searching and tracking (IRST) system demonstrates the advantages of high quality of image, strong resistance to electronic interference, good capacity of hiding, high performance-price ratio, and compactness of structure. Infrared detection and tracking system plays an important role in precision guidance, space-based earth observation, port and border monitoring, animal scientific research. Accurate infrared small dim target detection system contributes to improve the observation distance and tracking performance.In the concept of infrared small target, "small" not only means the small size of the target pixels, but also means the low contrast between targets and the background. Infrared small target detection is an extremely challenging task. In infrared image, clouds are brightness regions, which are usually brighter than the target regions, and the existence of dark current noise reduced the signal-noise ratio of image. These characters make it difficult to detect small targets. Additionally, the shape information and color information of the infrared small targets is absent, which further increase the difficulty of detection.In recent years, a large amount of methods have been proposed in the area of infrared small target detection in domestic and overseas. In this paper, two new methods for infrared small target detection have been proposed with extending or improving previous researches. The final method is the fusion of above two methods and the detection performance are significantly improved. The main contents and innovation of this paper is listed below:(1) Improved the conventional TDLMS filter and then proposed the bilateral TDLMS. In this method, the image is filtered by an improved TDLMS filter rightwards and leftwards respectively, then the fusion of the filtered images is used for detecting targets. The conventional TDLMS filter is improved by three points. The first one is that the infrared image is filtered by a Gaussian filter in order to reduce the noise. Experiments show that the Gaussian filter with the variance which is proportional to the noise variance contributes to good detection performance. The second on is that an adaptive method for iterating step selection is proposed. The last one is that the filtering window of TDLMS filter is down sampled from the conventional filter, which can reduce the computation. Comparative experiments prove that bilateral TDLMS outperforms the conventional TDLMS method and other improvements in terms of detection performance and computational efficiency. Further, a new noise estimation model is established in order to estimate the image noise variance quickly and accurately.(2) Proposed a novel detection method based on principal curvature. This principal curvature function method takes advantages of the principal curvatures of the image surface to detect targets, in detail, the positive Gaussian curvature and the negative mean curvature is used. In order to suppress noises of infrared images, a Gaussian filter is introduce. In order to optimize the detection parameters, noise, target and small size clouds are modeled by Gaussian distribution approximately. Experiences show that this method has better detection performance and computational efficiency than several popular methods.(3) Proposed a fusion method that fusion the detection results of bilateral TDLMS and principal curvature function method. The detection result of every pixel with the above two methods are classified by machine learning methods. Experiences show that the optimized parameters obtained by SVM with logarithm features own better performance than other methods. The detection results of the fusion model has better detection performance than the sole bilateral TDLMS and the principal curvature function method, especially for high noise and complex background images.