Infrared small target detection based on gray intensity descent and local gradient watershed

Robust infrared (IR) small target detection is an essential part of infrared search and tracking systems. Existing IR target detection methods based on human visual system usually use statistical features of rectangular region to measure the difference between target and background. In the case that the dim target size is smaller than the rectangular window cell size, since the window cell contains both the target and the background pixels, the statistical features cannot well represent the target contrast. The use of a rectangular window structure for calculation reduces the statistical difference between the target and the background, that is, reduces the saliency of the target, resulting in a slow rise in the ROC curve of the HVS method. This paper proposes a novel nonwindow structure filter for small target detection. Firstly, we obtain candidate target points by combining the high-frequency information of the image with the local maximum points, and at the same time, calculate the Sobel gradient map to prepare for subsequent processing. Then, a filter is designed based on gray intensity descent and local gradient watershed (GID-LGW) characteristics of the target, and it is used to calculate the contrast of each candidate target point. Finally, an adaptive threshold operation is applied to extract the target. Experiments show that compared with several baseline methods, the proposed method has a better detection rate, lower false alarm, and high speed, especially for small and weak targets with a size smaller than 3 x 2.

Automated summary

Robust infrared small target detection is crucial for infrared search and tracking systems. Existing methods typically utilize statistical features to measure the difference between target and background, which may not accurately represent the target contrast for small targets. This paper proposes a novel non-window structure filter for small target detection, combining high-frequency information and local maximum points to obtain candidate target points and calculating a Sobel gradient map for subsequent processing. The filter, based on gray intensity descent and local gradient watershed characteristics, is then used to calculate the contrast of each candidate target point. Experimental results demonstrate that the proposed method outperforms several baseline methods in terms of detection rate, false alarm rate, and speed, especially for small and weak targets with a size smaller than 3 x 2.