Wide Field-of-view Infrared Small Target Real-time Detection On Aerial Platforms

With the development of technology,infrared small target detection systems can be flexibly installed into a wide range of platforms.For wide field-of-view infrared systems on aerial moving platforms,infrared scenes exhibit complexity and change radically.Traditional target detection algorithms based on multiple frames or over-simplified scene assumptions are hardly applicable in such cases.Therefore,we address target detection with wide field-of-view on aerial moving platforms in this thesis.To overcome radical change of scenes,we develop algorithms tailored to detection in a single frame,and specifically algorithms to suppress two types of challenging background clutter – ground clutter and cloud clutter.To suppress ground clutter and cloud clutter,an algorithm based on local contrast measure is proposed.Firstly,local intensity peaks are detected as latent targets(or called region proposals)by background prediction based on modified top-hat transform.Then construct contrast measure NTHC with pixels in center region and neighbouring region of a latent target.Lastly use a classifier to predict class of latent targets from the feature.The proposed algorithm is robust to background clutter,sensitive to weak targets,and also simple to compute efficiently.Combined with the AREA feature,detection performance can be further improved.For higher computation efficiency,a decomposition algorithm is proposed to speed up morphological filtering with hollow structuring elements,which cannot be directly decomposed.The proposed algorithm decomposes filtering with hollow structuring elements into filtering with 1-D solid structuring elements.Combined with existing fast algorithms,at most 15 comparisons are needed to compute one pixel position,regardless of the size of structuring elements.Computational complexity is substantially decreased for large structuring elements.For systems with relatively more computation resources,where detection accuracy has higher priority than speed,an algorithm that combines target characteristics and background class prediction is proposed to further suppress ground clutter.The local background of a real target is sky,whereas that of ground clutter is ground.This difference can therefore be used to help classification.Specifically,contrast and geometry features are used to describe target characteristics,and KMFEAT,a high dimensional feature based on unsupervised learning,is used to describe local background of a latent target.Final detection is completed by combing scores computed from those features.Experimental results show that,compared with simple features,the proposed algorithm greatly improves accuracy.Similarly,an algorithm based on deep convolutional neural networks is proposed to further suppress cloud clutter.By training,a deep model can learn data representation automatically,without handcrafting.The proposed model has 9convolutional layers.It takes raw pixels as input,and outputs classification result after processing by the layers.Though the features are learning from data,high detection accuracy is achieved even for challenging cloud clutter.Finally,two approaches are proposed to combine the previous algorithms into a high detection performance infrared small target detection algorithm.Some of the test images are collected from sensors installed on aerial moving platforms.Experimental results show that,though the complexity of the test data,the combined algorithm can extract useful information from previous algorithms to achieve high detection accuracy.