Research On Target Detection Algorithms For Static Infrared Image

Infrared dim small target detection is an important research direction in infrared image processing,which acts an important role in first-detection and warning system,aircraft tracking and missile guidance.Due to the low SCR,small size and no obvious shape structure and texture information of infrared small target image,infrared small target detection is a very difficult task.Therefore,it is of great practical significance to realize the stable detection and tracking of infrared dim small target in complex background.It analyzes and summarizes the advantages and disadvantages of the existing infrared dim small target detection algorithm from many aspects.Aiming at the problem of infrared dim small target detection in different scenes,from the traditional image processing method to the deep learning method,the infrared dim small target detection algorithm is studied in a variety of ways,so as to improve the accuracy and robustness of the algorithm in different scenes.It includes:(1)Aiming at the problem that the existing image processing algorithms have poor accuracy in infrared image background prediction,an infrared dim small target detection algorithm,TLMS is proposed.Based on the angle of "main view" and MSE,the gray value of the central pixel is reconstructed from the pixels around the target.Then the predicted background image and difference image are obtained on the basis of the minimum mean square error criterion.Finally,we segment the infrared dim target by using adaptive threshold segmentation formula.Experiments show that TLMS method can suppress background and enhance infrared dim small targets more effectively.(2)Aiming at the problem that the performance of general target detection network and semantic segmentation network directly migrate to small target detection task is greatly reduced,it proposes a Gemini-Net based on semantic segmentation network model,which divides the whole task into two sub tasks,and the sub networks pay attention to their respective purpose(reduce the miss detection rate or false alarm rate)by designing the loss function to integrate the sub network information,the Gemini-Net model has an effective balance between the miss detection rate and false alarm rate.Experiments show that the Gemini-Net network is better than the general neural network model when it comes to the infrared dim target detection task.Apart from this,its performance is greater than the traditional image processing algorithm.(3)Aiming at the proportion of infrared dim small target in the whole infrared image,it can be regarded as "point target".Inspired by the idea of key points,it introduces the key point detection network to treat the infrared dim small target detection task,takes the semantic segmentation network as the backbone structure to extract the infrared image features,and uses the target center point coordinates for Gaussian calculation to generate the heatmap,which is compared with the ordinary direct image.The method of regression key points is more accurate than the method of heatmap regression key points.The infrared dim small target can be located by using the signal-to-noise ratio characteristics of the image and the Gaussian distribution characteristics of thermal map.Experiments are carried out on different datasets.The experimental results show that the key point detection network is functional in infrared dim small target detection task,which shows that the key point detection idea has certain scalability in many fields.