Research On Infrared Small Target Detection Methods Under Complex Background Based On Deep Learning

In recent years,the rapid development of various types of aircraft and navigators has posed serious threats to public safety,personal privacy,and public order.The detection of aircraft through infrared devices has been developed into a more commonly used detection method.In practical applications,the long imaging distance means long response time.But the detection target is usually in the form of small targets or even point targets in the farreaching infrared image.Therefore,it is essential to detect small targets quickly and reliably in infrared images.In the existing infrared small target detection technology,the human vision-based method is low cost and easy to realize.Consequently,research on human vision-based small target detection technology is of great significance for preventing aircraft threats.However,this task also faces many challenges,such as complex background,small pixel target and low SNR.To solve the low detection accuracy problem caused by these factors,this thesis mainly studies infrared small target detection technology in complex scenes,according to the following two aspects:First,this thesis proposes a small infrared target detection method based on convolutional neural network.This algorithm mainly utilizes the convolutional neural network to drive the infrared small target problem.First of all,to address the problem of insufficient data sources in infrared images,this thesis uses data arguement methods to expand the number of original experimental data which using artificial noise,artificial targets,and artificial infrared background overlayed to generate artificial images,reaching the data size that can drive neural network training.Then,the artificial target is paired with the artificial image generated by the overlayed image to train the neural network to obtain the detection model of the infrared small target.Then,the results of neural network detection can be further processed by traditional methods to enhance the detection effect.Finally,the likelyhood image is thresholded to get the final result.Compared with the existing small infrared target detection methods,the algorithm proposed mehod gains a higher accuracy rate.Meanwhile,this method also has some drawbacks.For example,the discrimination between the target and the background is blur,and a small part of the detection data produces detection results with a high false alarm rate.Except that,compared with traditional infrared small target detection methods,this method is better than most traditional infrared small target detection methods.Secondly,in order to solve the shortcomings in the first method and to improve the infrared small target detection performance,this thesis proposes to use gradients of image pixel and generative adversarial network to infrared small target detection.The method includes two branches.The first branch calculates the eight-neighbor gradient of the original image to obtain the saliency map of the infrared small target.The second branch uses the generative adversarial network to detect the small infrared target.The saliency map and the output result of the generative adversarial network are concatenated to supplement the missed targets and filter out the false targets to confirm the real small infrared targets.Then this thesis employs the optimum thresholding to get the final result.Considering the target in the original data image is small,the skip connection layers are used in the decoder of the generative adversarial network.With this operator,the information loss can be reduced significantly,and can be more suitable for infrared small target detection.In the post-processing stage,for continuously sampled infrared images,this thesis uses the nearest neighbor distance matching method to track small targets.Compared with the traditional infrared small target detection methods,this thesis can improve the accuracy of infrared small target detection and suppress false targets effectively.