A new point target detection method

This paper proposes a new point target detection method based on multiscale morphological filtering and local characteristic criterion. First 8-directions morphological Top-hat transform are used to detect all the possible targets with different scales. Next, the adaptive threshold is adopt to obtain the Region of Interest (RoI) of the target and improve signal to noise ratio (SNR). Secondly, we remove the remaining background edges according to the local characteristic criterion between background edges and point targets. Finally, we make use of the matching relationship of interframe to remove noise and obtain point target trajectory. And the point target is successfully detected out in motion control system. The results show that the proposed method has a good performance to suppress complex background and meet real-time requirement.

Recently, a lot of people use morphological filtering [16] to detect point target in complex background. Top-hat transform in morphological filtering is used to extract the similar structure in the image with a certain morphological structure element. The effect of background suppression for morphological filtering depends on the size and shape of structural element. The role of structural element in morphological operations is similar to the filtering window. The conventional Top-hat transform algorithm only uses a structural element, which ignoring the target details in different directions. Recently, some people use multiscale morphological algorithm to detect weak and small seismic wave. In this paper, we propose a new detection method based on multiscale morphological Top-hat transform and local characteristic criterion. The complex background and noise are eliminated clearly and the point target is detected out using this method. The results prove that our algorithm is effective and meets the requirements of the parallel properties needed for easy implementation in a real-time hardware system.

Multiscale morphological filtering
Aircraft with long distance from infrared detector in the image is rarely accounted for 1 pixel, more often spreading into horizontal or vertical direction of 2 pixels, or 3×3 pixels, as shown in Figure 1. It is because that long distance imaging produces optical diffraction, and point target will diffuse into Eyre spot, causing energy diffusion to neighboring pixels. The size and shape of point target in the motion are changing, and the strong mobility of point target is more likely to occur in the complex scenes with different scales. (1) We use the larger structural element than the target to make open operation, so as to obtain the image background. Then we subtract the background image from the original image, which is the Tophat transform.
The proposed multiscale Top-hat transform is defined as follows:  We construct a large filter with n ROI , and get the results of suspected point target images with different Morphological distribution shown as formula (12).
The constant false alarm rate (CFAR) is used to deal with the above ROI and we get the candidate targets. The threshold is calculated as: where  is the mean value of the image,  the standard deviation, and Th is the adaptive threshold, and k is a fixed constant, which can be regarded as the threshold of SNR. We make binaryzation to the image by Th . To show the results of the algorithm, we assign k is 1.5 according to the SNR of point target to be detected. The results of the above process may also remain residual strong fluctuation background edges and noise. We use the local characteristic criterion between background edges and point targets to remove the interference point.

Local characteristic criterion
The remaining points not only consist of point target, but also exist high-frequency background edges in the image. In order to detect out point target instead of the edges, we use the 4-directions weighted filter algorithm in 5×5 window centered at I(i, j), as shown in Figure 3. We define L m (m=1 to 4) as the direction vector consisting of four coordinates around I(i, j) in the mth direction, which are given in (5).  where w x,y is the weighted kernel to weight the absolute differences between t I(i+x, j+y) and I(i, j). We normally think that the closest four-neighbor pixels are more likely to the center pixel, so we assign the larger value 4 to them. We let w x,y =2 to the second closest pixels. And so on, we assign the small value 1 to the four far points whose coordinates x and y are both ±2. We combine L 1 to L 4 into a column vector L. The weighted kernel w x,y corresponding to elements of L are obtained as follows: We define a new variable named direction ratio (DR) to distinguish point target from background edges. The direction ratio is calculated as the maximum of Considering the difference of DR value, we can distinguish point target from edges by setting a threshold T, and the choosing threshold is slightly larger than 1. Thus, we obtain the point target detector as follows: if DR is less than T, it is a point target, otherwise it is an edge pixel. In order to visually display experimental results, the threshold T is set to 2 in this paper. There may also remain some noise in the detected points, while the position of noise is fixed or out of order. In the actual case, fixed-base detector acquisition device is easy to eliminate the noise according to the continuity of the moving point target, and we can obtain the target trajectory using multi frame accumulation. If we use the moving base detector to grab point target image, the speed of turntable is known always, and the fixed noise can be eliminated within 5 frames due to the high frame rate. The random flicker noise can't establish relationship in the interframe, so it's easy to remove it. The point target uniforms an approximate linear motion in a very short time, and the number of passed pixels in the interframe can be regarded as a fixed value. We can use Hough line detection method [17] to detect out the point target and obtain the target trajectory. Figure 4 (a) is an original image from infrared image sequence of clouds background, and the sequence contains 1000 frames. We use the proposed algorithm to detect point target in the sequence. Image processing finished in MATLAB R2014a, and PC configuration is i7-4790 CPU (3.60GHz). Figure 4 (b) is obtained by multiscale morphological transform and the adaptive threshold. The candidate points are mainly residual strong fluctuation background edges and noise, which will disappear after using the local characteristic criterion. There remain about 5 candidate points in each frame which marked in red as shown in Figure 4 (c). The trajectories of each point target are obtained by using the interframe correlation shown in Figure 4 (d). The results of image processing prove that the proposed algorithm is effective to suppress complex background. In the following we analyze the proposed algorithm through the specific data. The formula of calculating the SNR of point target in infrared image engineering is as follows:

Result analysis
where t  is gray mean of the target area, and b  is gray mean of the local background area, so the molecular is the absolute energy of point target.
b  is the standard deviation of the local background area. By calculating the SNR of point target in the whole moving phase, the SNR varies between 0.4-2.9, which is very low that the naked eye can't find the target in infrared image. In order to further measure the effectiveness of the proposed algorithm, we compare the propoed algorithm with common algorithms such as the max-median filtering, DoG space-scale, BM3D, and Gaussian Mixture Model (GMM). The target detection probability (P d ), false alarm probability (P fa ) and the running time of the algorithm are selected as the evaluation indexes of the results, which are defined as following:  Max-median filtering and DoG are simple and meet real-time requirement, but they are not good at background suppression, so the false alarm probability is a little high. BM3D and GMM are very effective for complex background suppression, but the algorithm is complex. The proposed algorithm has high detection probability and low false alarm rate. At the same time, the proposed algorithm has low complexity and short running time.

Conclusion
This paper proposes a new method of point target detection based on omnidirectional multiscale morphological filtering and local characteristic criterion. First we use the 8-directions 3×3 structural elements to detect all the possible targets with different scales. Next, the adaptive threshold is adopt to obtain the RoI of the target and improve SNR. And then, we use the local characteristic criterion to eliminate the residual strong fluctuation background edges. Finally, the trajectory of point target is obtained after using the interframe correlation. The experimental results show that the proposed algorithm is effective and has a good inhibition effect on the complex background, and the running time of the algorithm is short, which meets the requirements of real-time for engineering.