Infrared Small Target Detection Under Complex Background Based On Human Visual System

Infrared(IR) small target detection under complex background is a challeging task in IR image processing field, it is difficult to obtain high detection rate, low false alarm rate and high detection speed since large background area with high brightness and complex background edges are usually inevitable interferences during detection. IR small target detection algorithms based on human visual system(HVS) can eliminate large background area with high brightness by utilizing local contrast information but not brightness information to detect small targets. However, it should be noticed that not only real small target but also complex background edges have local contrast information, the difference between them is that real small target usually spread equally in all directions, while background edges usually spread along one direction in local. Current detection algorithms based on HVS can be divided into two kinds according to whether they ultilize directional information during contrast information calculation. The first kind of algorithms does not ultilize directional information, so they are usually sensitive to complex background edges and will present a high false alarm rate although they can achieve a fast detection speed. The second kind of algorithms, on the other hand, ultilize directional information and can better suppress complex background edges, however, they calculate multi-directional contrast information on pixel level, leading to a long time consuming.In this paper, to overcome the defects of known algorithms, we proposed a fast double-stage IR small target detection algorithm based on HVS. Multi-directional contrast information is used in the proposed algorithm to eliminate high brightness background and complex edges, which is helpful to achieve a high detection rate and a low false alarm rate. Besides, to solve the time consuming problem of current algorithms based on multi-directional contrast information, we divided the whole algorithm into two stages, i.e., the block-level fast search stage and the pixel-level identify stage, and in the block-level fast search stage, we divided the raw IR image into blocks and search candidate target regions on block-level, which can reduce processing time significantly.First, in the block-level fast search stage, the whole image is divided into blocks, and a new block-level multi-directional contrast information algorithm is proposed for each block, to detect some candidate target regions. Similar to other multi-directional contrast information algorithms, this algorithm can eliminate both of high brightness background and complex background edges, which is helpful to achieve a high detection rate and a low false alarm rate. Besides, this algorithm is calculated on block level and data to be processed is significantly reduced, which is helpful to achieve a fast detection speed.Second, in the pixel-level identify stage, a new pixel-level convolution algorithm is proposed to obtain the target center in the candidate target regions. Current pixel-level convolution algorithms are not sensitive to directions and cannot identify real targets from complex background edges. The proposed algorithm, by utilizing a new kind of semi-ellipse filter templates, is sensitive to directions and can identify real targets from complex background edges, so it can further eliminate residual complex backgrounds in the fast search stage, and then achieve a better detection performance.Experimental results and comparison with other algorithms show that the proposed algorithm can achieve the lowest false alarm rate under the same detection rate, its TPR is larger than 95% while its FPR is less than 0.3% in all of the five IR sequences, which means that the average false target number is less than 0.7 in a frame. Besides, the average time consuming for an IR image with resolution of 320×256 is only about 0.025 s, which means that 40 frames can be processed in one second, better than current HVS IR small target detection algorithms based on multi-directional local contrast information.The research in this paper will be helpful to achieve IR small target detection with high detection rate, low false alarm rate and fast speed, and will be very useful for many fields involving IR detectors, such as IR guidance, IR early warning, ect.