Research On Infrared Image Detail Enhancement And Small Target Detection

Infrared imaging technology has been widely used in defense,industry,medical and other fields.A new generation of high-resolution,high frame rate and high dynamic range infrared imaging systems have been widely used in various applications with the rapid development of infrared detector technology,meaning that higher requirements are put forward for the performance and computational efficiency of infrared image processing methods.On one hand,for airborne,missile-borne and other applications with limited computing resources,a large number of conventional image detail enhancement and small target detection algorithms cannot meet the requirements the performance and computational efficiency.On the other hand,the improvement of the algorithm performance in complex backgrounds with powerful ground station computational resources remains to be further research.To meet the requirement of the development new generation of infrared imaging systems,this dissertation was engaged in infrared image detail enhancement and small infrared target detection algorithm.The main contents are described as follows:Firstly,a real-time infrared image enhancement method based on fast guided filtering and platform histogram has been proposed.The input image has been divided into the base and detail layer by cascading the fast guided filter and the Gaussian filter,the base layer and the detail layer has been used for gradient distribution and detail enhancement respectively,with detail enhancement and noise as well as gradient reverse suppression simultaneously.The enhancement results in different scenarios demonstrated that the proposed method yields a good balance between detail enhancement and noise and gradient reverse suppression,with 3% improvement compare with BF&DDE in terms of the mean opinion score(MOS)metrics.Moreover,the proposed method not only working robust on noisy image,but also can be simply accelerated by parallel processing and pipeline architecture,which means a promising application potential.Secondly,A small infrared target detection method based on fast preselection and local feature fusion has been proposed.To meet the requirements of the small infrared target detection with sky background,a fast detection method has been proposed according to the blob-like characteristics of the small target,and the difference of Gaussian filter and the cumulative distribution function were introduced to generate the binarized mask.The candidate pixels,which account for a proportion of 0.2% has been further processed in terms of the local contrast and gradient with an iterative segmentation.Experimental results demonstrated that the proposed method yield robust performance on images with sky background.The average area under the precision-recall(AUC)on five high-resolution challenging test sequences shown that the proposed approach achieving 0.87 and outperforms the FKRW method by 28.8%,with the running time on different datasets are less than the PSTNN with a margin of at least 26.9%.Thirdly,an infrared small target detection method based on centroid positioning residual network has been proposed.A small dataset for infrared target detection with the semantic segmentation annotation has been collected by using self-developed midwave and long-wave infrared cameras.To balance the conflicts between the target scale and network stacking,a combination of the Res Net-50 and the feature pyramid networks has been developed as the feature extraction network.To address the problem of the bounding box cannot be used directly in the post-detection tasks,an infrared small target centroid positioning residual network with feature embedding module has been proposed to output the target centroid directly.The primary contributions including output the target centroid directly with advantages on the method of detection,segment,and calculate centroids in sequence,and the metrics development of the mean localization average precision(m Loc AP).The experimental results shown that the proposed target centroid positioning residual network yields the best m Loc AP on test images with different scenarios,beating the ACM-Net with a margin of 12.7%,which further illustrates the practicality of the proposed method.