【作者】 李鹏；

【导师】 王杰；

【作者基本信息】 郑州大学 ， 控制工程（专业学位）， 2018， 硕士

【摘要】 随着科技的进步,人工智能逐渐应用于各种场合,特别是在图像处理方向,人工智能更是取得了巨大的进步。红外图像是通过热感应摄像机拍摄到的一系列灰度值图像,目标的热度越高,其在图像中显示的灰度值也就越大。通常,飞机、导弹等热量极高的物体通常在红外图像中显示为一个亮点,而云背景杂波的灰度值相对较低,因此可以清晰的看出飞机和导弹的运动迹象。但是,在目标距离摄像机较远时,或目标与云背景有重叠时,其灰度值并不比云背景高出太多,且目标的形状较小,通过人眼很难注意到。而且,人工观察红外图像的小目标需要消耗极大的人力,也无法做到24小时实时监控,准确性与实时性都不能够令人满意。因此,红外图像中弱小目标的自动检测技术应时而生,越来越多的方法被用来检测小目标。目前图像识别领域最好的算法通常是深度学习算法,如卷积神经网络(Convolutional Neural Networks CNN)、长短时记忆网络(Long Short Term Memory Network,LSTM)等。但深度学习算法通常无法用于识别红外图像的小目标,因为小目标在整幅图像中所占面积极小,对整幅图像的影响也不大。而深度学习算法通常是将整幅图像作为训练数据输入,因此直接采用深度学习算法很难准确地识别出小目标,而且深度学习还需要消耗大量的时间,不能够满足小目标识别的实时性要求。传统的小目标识别算法通常是基于提升小目标区域的对比度,之后通过一些图像分割算法,如阈值分割等,对小目标区域进行提取。采用这种方法可以应用于大多数情况,但当小目标的灰度值低于云杂波的灰度值时,传统的方法通常很难正常识别。为解决以上的问题,本文提出了一种将传统图像对比提升算法和机器学习算法相结合的小目标识别方案,同时包含了两类算法的优势,在保证小目标识别准确率的同时,极大地降低了小目标识别的虚警率,即很少会错误检测,把背景识别出小目标。本文算法分为三步:首先,采用LoG卷积函数对原始包含小目标的红外图像进行卷积,提升小目标区域的对比度;其次,采用阈值分割算法,设定一个较低的阈值,提取出所有可能是小目标的11*11像素区域;最后,采用核支持向量机(Kernel Support Vector Machine,KSVM)对上一步提取的区域进行识别,最终确定小目标的位置。本文算法避免了将大量无用背景数据输入KSVM,从而极大的提升了算法的运行速度。而KSVM在最后一步确定小目标区域时,可以极大地降低虚警率,提升算法的识别效果。通过四组红外图像的实验验证,本文算法在保证正确率的同时,虚警率远低于传统算法。更多还原

【Abstract】 With the advancement of science and technology,artificial intelligence is gradually applied to various occasions,especially in the direction of image processing,and it has made great progress.An infrared image is a series of grayscale images captured by a thermal sensor camera.The higher the heat of the target,the greater the grayscale value it displays in infrared images.In general,targets with extremely high heat,such as planes and missiles,usually appear as bright spots in infrared images,while the gray value of clutter on the cloud background is relatively low.Thus the signs of movement of planes and missiles can be clearly seen.However,when the target is far from the camera,or when the target overlaps with the cloud background,its gray value is not much higher than the cloud background,and the shape of the target is small,which is difficult to notice by the human eye.Moreover,manually observing the small target of the infrared image requires a great deal of manpower,and it cannot achieve 24-hour real-time monitoring.The accuracy and real-time performance are not satisfactory.Therefore,the automatic detection technologies of dim and small targets in infrared images emerge from time to time,and more and more methods are used to detect small targets.The best algorithms in image recognition are usually deep learning algorithms such as convolutional neural networks(CNN),long-short memory networks(LSTM),and so on.However,the deep learning algorithms can not usually be used to identify the small target of the infrared image,because the small target only occupies a very small area in the entire image and has little effect on the entire image.But the deep learning algorithms usually use the entire image as input for training data.Therefore,it is difficult to accurately identify the small target directly using the deep learning algorithm.In addition,the deep learning algorithms also requires a lot of time,then can not meet the real-time requirements of the small target recognition.In the other hand,the traditional small target recognition algorithms are usually based on the enhancement of the contrast of the small target area.After that,some image segmentation algorithms such as threshold segmentation are used to extract the small target area.This method can be applied to most cases,but when the gray value of the small target is lower than the gray value of the cloud clutter,the conventional method is usually difficult to identify correctly.In order to solve the above problems,this paper proposes a small target recognition method that combines the traditional image enhancement algorithm and machine learning algorithm.It includes the advantages of the two types of algorithms.While ensuring the accuracy of small target recognition,the false alarm rate of small target recognition is reduced.That means the cloud background is rarely recognized as a small target.The algorithm of this paper is divided into three steps: First,the LoG convolution function is used to convolute the infrared image that originally contains the small target,to enhance the contrast of the small target region;Second,the threshold segmentation algorithm is used to extract some 11*11 pixel regions that may be small targets,which needs to set a lower threshold;Third,use kernel support vector machine(KSVM)to identify the regions extracted in the previous step,and finally determine the location of the small target.This algorithm avoids inputting large amount of useless background data into KSVM,which greatly improves the running speed of the algorithm.When KSVM determines the small target area in the last step,it can greatly reduce the false alarm rate and improve the recognition performance of the algorithm.The experimental results of the four groups of infrared images show that the false alarm rate is far lower than the traditional algorithm while the algorithm guarantees the correct rate.更多还原