【作者】 王昕；

【导师】 张佳岩；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2025， 硕士

【摘要】 目前,我国卫星技术高速发展,已跻身于全球领先行列,遥感卫星由于自身的特殊应用,成为不可或缺的一部分。遥感可见光图像目标检测作为其中的一项关键技术,在人们日常生活、军事侦察、城市规划等领域均有着广泛的应用。但是,受制于成像距离、成像角度等因素的影响,遥感可见光图像中目标的尺寸通常较小、尺度多样、特征不明显且易受到干扰,导致现有方法的检测性能不够优异,无法高精度地检测出感兴趣的目标。此外,由于遥感可见光图像在成像时会受到环境因素和载荷工作条件的干扰,目标的特征会发生遮挡、模糊等退化情况,进一步影响了检测的效果。最后,受制于传输时延和传输带宽,对大数据量的遥感可见光图像进行星上处理成为主流选择,考虑到功耗、抗辐射、可靠性等问题,星上多以边缘设备为主,有限的资源也会限制算法的设计。因此,本文针对上述存在的问题,展开高精度、低复杂度、泛化能力强、适用于边缘设备的目标检测算法的研究,并进行验证。主要的工作内容如下:首先,针对遥感可见光图像存在的目标尺寸小、细节信息少、背景中场景复杂引起的漏检和虚警情况,在保证轻量化的前提下,优化了特征的提取、设计了自适应的特征融合以及基于维度交互的特征增强,提升目标检测算法的性能。最后以本文的数据集为基础,从算法的复杂度、参数量、准确率等角度,验证所设计的算法的性能,实验表明,所设计的算法优于当下的主流轻量化目标检测算法。其次,考虑到遥感可见光图像在成像过程受到成像环境、成像设备工作条件等因素的影响,导致目标检测算法性能的劣化,针对这一问题,采用深度学习与大气散射模型、锐化滤波相结合的方式,设计基于自适应增强的退化遥感可见光图像的目标检测算法,在原有设计算法的基础上,略微增加复杂度,解决由于退化引起的检测精度下降的问题。最后,基于ZYNQ平台的特点,进行算法部署的整体框架设计,将算法分为PL和PS两个部分实现。其中,PL端主要完成图像数据的计算,利用HLS编写了算法中的卷积、全局池化、Concat、上采样IP核,并进行适度优化,搭建了硬件电路,PS端完成了整体流程的控制、数据的读取与写入以及后处理操作。最后,在ZYNQ平台上完成了所设计算法的部署,经过实验验证,基于ZYNQ平台上的目标检测算法可以完成目标的准确检出。更多还原

【Abstract】 At present,China’s satellite technology is developing rapidly and has become a global leader.Remote sensing satellites have become an indispensable part due to their special applications.Remote sensing visible light image object detection,as a key technology,has a wide range of applications in people’s daily lives,military reconnaissance,urban planning,and other fields.However,due to factors such as imaging distance and angle,the size of targets in remote sensing visible light images is usually small,the features are not obvious,and they are easily affected by interference,resulting in insufficient detection performance of existing methods and inability to accurately detect the target of interest.In addition,due to the interference of environmental factors and payload working conditions during the imaging of remote sensing visible light images,the characteristics of the target may undergo degradation such as occlusion and blurring,further affecting the detection effect.Finally,due to transmission latency and bandwidth constraints,on-board processing of large amounts of remote sensing visible light images has become the mainstream choice.Considering issues such as power consumption,radiation resistance,and reliability,edge devices are mostly used on board,and limited resources also limit algorithm design.Therefore,this article focuses on the above-mentioned problems and conducts research on high-precision,low complexity,strong generalization ability,and edge device applicable object detection algorithms,and verifies them.The main job responsibilities are as follows:Firstly,in order to address the issues of missed detection and false alarms caused by small target size,limited detail information,and complex background scenes in remote sensing visible light images,while ensuring lightweight,we optimized feature extraction,designed adaptive feature fusion,and feature enhancement based on dimensional interaction to improve the performance of object detection algorithms.Finally,based on the dataset presented in this article,the performance of the designed algorithm was validated from the perspectives of algorithm complexity,parameter count,and accuracy.The experiments showed that the designed algorithm outperformed the current mainstream lightweight object detection algorithms.Secondly,considering that remote sensing visible light images are affected by factors such as imaging environment and imaging equipment working conditions during the imaging process,which leads to the degradation of target detection algorithm performance,a target detection algorithm based on adaptive enhancement of degraded remote sensing visible light images is designed by combining deep learning with atmospheric scattering models and sharpening filtering.On the basis of the original design algorithm,the complexity is slightly increased to solve the problem of decreased detection accuracy caused by degradation.Finally,based on the characteristics of the ZYNQ platform,the overall framework design for algorithm deployment is carried out,and the algorithm is implemented in two parts:PL and PS.Among them,the PL end mainly completes the calculation of image data,and uses HLS to write convolution,global pooling、Concat、Upsampling the IP core and conducting moderate optimization,the hardware circuit was built,and the PS end completed the overall process control,data reading and writing,and post-processing operations.Finally,the deployment of the designed algorithm was completed on the ZYNQ platform,and after experimental verification,the target detection algorithm based on the ZYNQ platform can accurately detect targets.更多还原