| Due to the characteristics of infrared imaging systems,the generated images often have background areas with variable brightness and shapes,as well as noise caused by various factors.The small target signal is weak,except the small target size,there is a lack of shape characteristics,texture information,etc.,making the detection of small targets in the infrared search and tracking（IRST）system a difficult task.Therefore,how to enhance the small target information and suppress the interference effect of background noise is of great significance for IRST system.In IRST system,the key points and challenges of small target detection task are as follows:how to effectively analyze and describe the internal features of small targets,distinguish targets from backgrounds by constructing feature expressions,and design algorithms that can better adapt to complex environments and have the ability to locate and recognize small targets.Considering the above problems,this thesis mainly focuses on the characteristics of Human Visual System（HVS）,from the analysis and expression of the significant features between the target and the background in the infrared small target image carried out research,and proposed the algorithm that can complete the infrared search and tracking system on the small target detection task.The main innovation points and research contents of this thesis are summarized as follows（1）Aiming at the problem of weak target signal strength in infrared small target images leading to the difficulty of detection,a weighted accommodative scale local contrast calculation model based on human visual contrast characteristics is proposed.By studying the HVS theory,it can be known that the perception of the target by human vision mainly comes from the contrast difference between the target and its surrounding background,not just the signal strength of the target.By calculating and enhancing the contrast difference between the target and its surrounding background,the adjustable local contrast measurement model generates a saliency map between the target area and its adjacent background,which is suitable for small targets with different scale changes in different backgrounds by emphasizing the target area,and limiting the background interference area.In addition,the target enhancement method based on the signal-to-clutter ratio is introduced to further enhance the target information.The experimental results show that this method has better ability to adapt to different background images.Compared with the multiscale relative local contrast method（MRLCM）in the experimental analysis,the signal to clutter ratio gain value of the proposed method is about 28%higher.（2）Aiming at the problem of poor real-time detection due to the high complexity of model,an infrared small target detection algorithm based on multi-scale and multi-directional saliency map is proposed.According to research and analysis,it is known that,on the one hand,the gray value of pixels in the target region and its neighborhood region has no correlation and discontinuity,and the difference between the gray value and its neighborhood region has no direction.On the other hand,the background region is mostly in the flat and uniform region,and the edge region in particular is different from its neighbors in some direction.Based on this analysis,the multi-scale structure and the multi-direction difference model are designed by analyzing the multi-scale and multi-direction saliency maps in the global image,which are used to calculate a variety of saliency maps in different scales and different directions.Then,by extracting and selecting the correlation saliency maps,and the final fused saliency map is generated,which makes the saliency of small target more prominent in the detection of infrared small targets,while weakening the background information.Compared with the method of chapter 1,this method significantly reduces the computational complexity of the algorithm.Moreover,compared with the multiscale patch based contrast measure（MPCM）,under the same false detection rate of 2.5×10-5,the method improved the detection rate by about 10%on average.（3）Aiming at the problem of false detection due to interference caused by clutter signals similar to small targets in complex background region,this thesis proposes a probability graph calculation model based on random walk for infrared small target detection and divides the algorithm process into selecting the region of interest and analyzing the region of interest.By analyzing the distribution characteristics of gray pixel values in the background region and the target region in the whole image,the background estimation model is constructed,so that most of the background regions can be estimated and suppressed in the original image.At the same time,according to the difference between the target region and the neighborhood region,the salient feature region is further highlighted,and the pixel of the candidate region of interest is obtained.Then,probabilistic model is used to calculate the saliency value based on probability for the selected pixels in the region of interest.Finally,the selected pixels with probability saliency feature description are enhanced by saliency image of first step so that the actual target pixels can be better segmented from the complex background region.The experimental results show that,according to the receiver operating characteristic curve of the evaluation index,the proposed method reduces the false detection of pixels by about 35%compared with the facet kernel and random walker（FKRW）method when the detection rate is 80%.（4）Aiming at the problem that weak infrared small target signals are easy to be missed leading to low detection rate and poor generalization ability in complex scenes,a saliency fusion model based on local differential and convolutional networks is proposed for infrared small target detection.Because deep learning network can learn more image features and has better adaptability,this thesis applies the deep learning segmented network to the detection of small targets in infrared images.On the one hand,by combining HVS characteristics and attention module,a network model is constructed in this thesis,which can learn to segment the target pixels in the image by deep convolutional network,so as to obtain the segmented saliency map.On the other hand,this thesis also calculates the local difference between the object and the neighborhood,and obtains the saliency map of pixels represented by the difference.Finally,two different saliency maps are fused to further suppress false alarm pixels in the background region,so that the saliency map could be used to segment the target from the background region to achieve the purpose of target detection.It can be seen from the experimental results that this method can highlight the infrared small target information in the background image,and the performance of detection accuracy is good in the experimental analysis.Under the premise of the same false detection rate of2.5×10-5,the detection rate of the proposed method can be improved by about 3%on average compared with multiple feature representation（MFR）methods based on the deep learning framework.In this thesis,four infrared small target detection methods based on the characteristics of human visual system are proposed.The presented target saliency feature extraction methods have good performances in improving the target detection rate and reducing the false detection rate.At the same time,the methods can effectively enhance the signal strength of small targets in infrared images and resist noise interference.The task of infrared small target detection can be effectively completed.They can be widely used in infrared search and tracking related visual tasks,and have good generalization ability and practicality. |
PDF Full Text Request