Research On Techniques Of Infrared Target Detection Based On Feature Mining And Appearance Modelling

Infrared target detection technique uses passive imaging to capture targets,and has been extensively applied in both military and civil areas,such as guidance,early warning and security,owing to its advantages of high concealment,strong ability of anti-electromagnetic interference and so on.With the development of computer vision,using feature mining and appearance modelling to make descriptions of target of interest and employing image understanding and analysis tools to develop intelligent target detection and tracking algorithms have become the inexorable trend of infrared target detection technique.Affected by the challenges generated from variety of target type,background complexity and noise interference,almost all the current infrared target detection algorithms suffer from low detection precision and weak robustness.This paper focuses on the following key issues: infrared image enhancement,infrared image segmentation,infrared small target detection and infrared target tracking,and designs a series of image analysis and processing algorithms which can meet different detecting requirements by mining deep image features and building effective appearance models of infrared target so as to provide theoretical supports for the construction of infrared target detection system.The main works of this paper are as follows:In order to preserve the details of image and avoid the over-enhancement of background and noise pixels,we propose a local entropy weighted feature histogram equalization-based infrared image enhancement algorithm,which establishes the detail feature distribution of infrared image and separately equalizes the foreground and background feature histograms by setting adaptive double plateau thresholds.Besides,we propose an infrared image enhancement algorithm based on adaptive partition of grayness interval and brightness correction.Local entropy weighted feature enhancement and visual factor-based correction are made for the foreground intervals while linear gray level mapping is implemented for the background intervals.Moreover,the whole brightness of output image is corrected using a reference image.This algorithm improves the local contrast of infrared image and weakens the poor visual effect caused by gray level merge,noise amplification and brightness distortion.Since traditional active contour models cannot deal with infrared images with disadvantages,like intensity inhomogeneity and blurred boundary,we present a multi-featurebased geometric active contour model that improves the level set equalization by using an adaptive weighted coefficient matrix to combine the average intensity feature-based global signed pressure function and the multi-feature-based local signed pressure function.By this means,its representation ability of local appearance is increased and can avoid the phenomenon of contour ‘leak'.What is more,the performance of image segmentation is almost not influenced by the initial state of level set function.Considering that conventional spatial filtering-based algorithms are not able to effectively suppress false alarms,we first present an over-complete dictionary learning-based single-frame infrared small target detection algorithm.It uses the sparsity of the representation coefficient of candidate patch over the small target dictionary to model the target appearance and select the real target using the proposed index called sparse rate.Owing to this kind of processing,the single detecting false alarms are significantly suppressed while the detection rate is kept.In addition,most of the existing multi-frame infrared small target detection algorithms only utilize the spatial correlation of target positions,making their inter-frame detection accuracy always limited.Thus,we propose a saliency histogram and graph matching error-based multi-frame detection algorithm.Based on the visual saliency analysis made in spatial domain,it further distinguishes the real target and false alarms according to the graph matching error calculated by the relative positions between candidate targets in temporal domain.This algorithm can not only extract bright and dark targets at the same time,but can also increase the precision of multi-frame detection remarkably.In order to overcome the sensitivity to infrared noise,a total variation constraint-based infrared target tracking algorithm is designed.It decomposes the appearance model of candidate target into a target term with low-rank property and an occlusion term with sparse property,and prevents the noise pixels from being judged as occlusion by imposing total variation constraint over the occlusion term.To this end,the tracking drift problem caused by noise interference is effectively solved.Meanwhile,a temporal consistency and contiguous occlusion constraintbased infrared target tracking algorithm is developed to reduce the influence of target appearance change and occlusion interference.On the one hand,the appearance similarity in adjacent frames is controlled by imposing temporal consistency constraint on the coefficient vector of dictionary coding;then,the shape distribution of occlusion is controlled by modelling a binary occlusion vector with contiguity constraint.This algorithm increases the tracking robustness to pose variation,environmental illumination change and partial occlusion.