Algorithm Research On Infrared Small Target Detection

The detection of infrared small target has been a very hot subject in current research in the military field, which is related to our military force. At the same time, in the medical, industrial and transportation and other fields, infrared dim target detection has wide application value. While the high altitude small target detection become a difficulty and a focus of detection research because the target has weak energy and it’s imaging area is very small.This article first analyzes the three main components about the feature of infrared image including the small target, background and noise. Then we establish the small target and background model from three aspects of time domain, spatial domain and time and space domain, and analyze several common noise in the image. On this basis, we introduce several typical image preprocessing algorithm, which is aim to suppress noise, and improve the signal-to-noise ratio of image. Through simulation experiments, we found those basic algorithms can meet the demand, but there is a lot of noise point. To solve the problem, this paper proposes a hybrid filter, which is constructed based on the principle that every filter will not make the same mistakes. The experimental results show that this filter is better than other traditional filter.It’s not easy to detect the infrared target because the target has weak energy and small size. This paper presents a method of energy accumulation based on the time and space domain. In this way, the image signal-to-noise ratio has been improved. Because the traditional single frame detection has many false alarms, a new detection method based on the image sequences for moving dim small targets is presented. Since the target in the time domain has a certain motion, the projection onto a 2D plane will be a continuous trajectory. While the trajectory image may cause local fracture because of fast moving. To avoid this situation, we project the trajectory after appropriate expansion, and detect the target location according to the trajectory. Because of the immobility of the dead point noise, we further remove the false alarms by increasing the jump counter.This paper introduces the classical classification algorithm of AdaBoost from the view of pattern recognition, and regards the target detection as two classification problem, and proposes an improved classification algorithm. The main idea is that we do not train each dimension feature when training the stump weak classifier, we just directly train the weak classifier through a regression method. The contrast experiment with other deformation algorithm of AdaBoost shows that the classifier has better effect, and the calculation speed is faster.