【作者】 张立材；

【导师】 刘沛津；

【作者基本信息】 西安建筑科技大学 ， 电气工程， 2022， 硕士

【摘要】 图像融合技术是通过把各种传感器采集到的同一目标的图像信号加以融合,以增加对场景的感知和对目标的识别能力等。红外与可见光图像融合是图像融合领域中的重要内容,它是通过整合红外与可见光图像中的优势信号,得到符合实际需求的信息量丰富的清晰图像,被广泛地应用于计算机视觉、目标探测与识别、视频监控、军事等诸多领域。针对目前的融合算法还存在源图像细节信息保留不足、红外目标突出能力不足、融合中引入较多伪影、算法消耗时间较长等问题,本文通过对图像融合方法的理论研究,兼顾融合质量和效率的前提下,提出了两种新的多尺度融合方法,分别是:(1)凸显细节的红外与可见光图像融合方法;(2)提高实时性并凸显细节的红外与可见光图像融合方法。(1)为提升细节信息保留能力、减少伪影引入、突出红外目标,提出第一种方法。首先,对可见光图像采用一种基于引导滤波器结合动态范围压缩和对比度恢复的方法进行图像对比度的增强。其次,通过使用基于引导滤波器的图像多尺度分解来获得红外与可见光源图像的基础层和细节层。对于基础层融合,提出了一种基于对源图像的细节和能量测量来确定图像的像素值,这样可以减少融合的能量损失和纹理细节特征突出显示以获取更多的源图像细节。然后,为了融合细节层,采用多尺度梯度函数生成显著性映射,归一化后,从它得到一个权重映射,这样算法的复杂性可以大大减小了。最后,采用递归分离和加权直方图均衡方法对融合后的图像进行优化。(2)在前一种方法的基础上,为减少算法消耗时间,提出第二种方法。首先采用基于引导滤波的动态范围压缩和对比度恢复相结合的方法,提高可见光源图像的对比度。然后,使用基于引导滤波的图像多尺度分解,从而获得红外与可见光源图像的基础层和细节层。在基础层中,采用加权平均的融合规则,减小算法复杂度。针对主要耗时的细节层融合,提出一种新的基于视觉显著性的权重构建方法,通过对源图像取引导滤波和中值滤波的差来计算显著图,并将显著图归一化后得到权重图,以此对细节层进行融合,大幅减小算法的复杂度的同时,可以保留源图像的细节信息。最后,通过对红外与可见光图像的融合实验表明,本文所提的第一种方法,能够在保证融合图像良好的视觉效果的同时得到能够真实清晰地反映目标信息和场景细节信息的融合图像,在此基础上提出的第二种方法可以兼顾保留源图像细节信息与缩短算法消耗时间,因此本文所提出方法可为视频实时监控、无人驾驶等领域提供方法基础。更多还原

【Abstract】 Image fusion technology is to fuse the image signals of the same target collected by various sensors to increase the perception of the scene and the ability to recognize the target.Infrared and visible image fusion is an important content in the field of image fusion,which integrates the dominant signals in infrared and visible images to obtain clear images with rich information that meet actual needs.It is widely used in computer vision,target detection and recognition,video surveillance,military and many other fields.For the current fusion algorithm,there are still problems such as insufficient source image detail information retention,insufficient infrared target highlighting ability,more artifacts introduced in fusion,and long algorithm consumption time.In this paper,two new multi-scale fusion methods are proposed through theoretical research on image fusion methods,taking into account the fusion quality and efficiency.They are:(1)an efficient algorithm to highlight details in infrared and visible image fusion.(2)an efficient algorithm to improve the real-time performance and highlight details in infrared and visible image fusion.(1)In order to improve the ability to retain detailed information,reduce the introduction of artifacts,and highlight the infrared target,the first method is proposed.First,a method based on guided filter combined with dynamic range compression and contrast restoration is used to enhance the image contrast of visible images.Second,the image is decomposed into base layer and detail layer by a multiscale decomposition method based on guided filter.For base layer fusion,a method is proposed to determine the pixel value of the image based on the detail and energy measurement of the source image,which can reduce the energy loss of fusion and highlight texture detail features to obtain more source image detail.Then,in order to fuse the detail layer,a multi-scale gradient function is used to generate a saliency map,and after normalization,a weight map is obtained from it,so that the complexity of the algorithm can be greatly improved and reduced.Finally,the fused images are optimized using recursive separation and weighted histogram equalization.(2)On the basis of the former method,in order to reduce the time consumption of the algorithm,the second method is proposed.First,a method based on guided filter combined with dynamic range compression and contrast restoration is used to enhance the image contrast of visible images.Second,the image is decomposed into base layer and detail layer by a multiscale decomposition method based on guided filter.In the base layer,the weighted average fusion rule is adopted to reduce the complexity of the algorithm.For the main time-consuming detail layer fusion.A new weight construction method based on visual saliency is proposed.The saliency map is calculated by taking the difference of guided filtering and median filtering on the source image,and the weight map is obtained after normalizing the saliency map,so as to fuse the detail layer,which can greatly reduce the complexity of the algorithm and at the same time retain the detail information of the source image.Finally,the fusion experiment of infrared and visible light images shows that the first method proposed in this paper can obtain a fusion image that can truly and clearly reflect the target information and scene detail information while ensuring the good visual effect of the fusion image.On this basis,the second method can well reduce the time consumption of the algorithm while retaining the details of the source image,so it can provide a method basis for real-time video surveillance,unmanned driving and other fields.更多还原