Fast capture of appointed infrared targets based on estimation of rotation angle

For the problem of recognizing infrared targets, a method based on estimation of rotation angle was proposed. The method first segmented the image by local adaptive threshold and mark connected areas. Integral image was used to accelerate the computation of the local threshold. The connected areas were resized to the same size. The pixel values of the resized images were used as features. Then the rotation angle of the target was estimated by a neural network. Some areas were filtered during the estimation. At last, the area was recognized by another neural network. For some applications, the number of sample was not sufficient and a little difference existed between the samples and targets. For this problem, a method based on random growth and erosion was proposed to generate samples. Experiments show the method is effective and has a high recognizing rate even when the shape of samples is not exact.