高分辨率空间CMOS相机视频实时压缩技术研究

发布时间：2018-05-12 10:23

本文选题：CMOS相机 + Karhunen-Loeve变换　；参考：《中国科学院研究生院(长春光学精密机械与物理研究所)》2016年博士论文

【摘要】：由于空间相机在视频分辨率以及帧频等指标要求的不断提高,导致CMOS输出的视频数据量急剧增加。现有的压缩算法运算过程十分复杂,而且主要停留在软件仿真阶段,硬件实现困难且效果较差。为了有效缓解星上存储器以及信道传输的压力,研制出压缩性能好、数据处理快的实时视频压缩系统迫在眉睫。论文以参与的工程项目“空间CMOS相机技术”项目为背景,对全色和多谱段CMOS相机视频压缩技术分别进行了研究。现将本文的主要研究内容和成果概括如下：1、根据项目需求以及CMOS视频的特点,探讨小型CMOS相机及高分辨率CMOS相机ASIC的ADV212压缩技术。提出利用Custom-specific工作模式为各种格式的视频提供接口,并通过FPGA内部的块RAM以及DDR3 SDRAM的乒乓操作对数据进行缓存,显著地提高了工作效率;之后,为了适用于不同应用场合,本文方法实现了码流的存储后传输以及直接传输之间的切换,并通过纠错编码极大的提升了闪存的纠错能力；最后,为了验证方法的可行性,本文基于压缩板以及解压板进行了实验验证。结果表明,压缩系统可实现实时稳定的工作,通过软件设置,系统可以实现极高的压缩比,压缩比80：1时,平均峰值信噪比(PSNR)高于28 dB,压缩比150：1时,平均PSNR高于26 dB,解决了在大数据量下,压缩系统硬件实现困难以及实时传输困难等问题。2、结合离散小波变换(Discrete Wavelet Transform, DWT)与Karhunen-Loeve变换(KLT)提出一种空间多光谱视频压缩算法。通过两种变换的有效结合可以将图像的能量集中到少数系数上,更好地达到压缩效果。本文首先将多光谱图像的每个谱段进行快速2维离散5/3小波变换,消除多光谱图像的大部分空间冗余。然后对所有谱段产生的小波系数进行改进KL变换,来消除光谱冗余和残存的空间冗余。最后对所得系数进行熵编码,得到压缩码流。实验结果表明,在0.25-2bit/pixel (b/p)范围内,平均PSNR达到41dB,与其他多光谱图像压缩算法相比,极大的提高了系统PSNR,提升了多光谱图像压缩算法的性能；同时提出了硬件实现策略,验证了本文理论的正确性以及算法的可行性,为空间多光谱图像压缩系统实现提供了参考。3、图像进行KLT和一级DWT后,不同谱段的系数之间以及同一谱段的高频子带之间仍然存在很大的相关性,其平均值大于0.9,提高小波变换级数后此值会相应的降低,但效果不是十分理想；基于DWT和Tucker分解的压缩算法将图像作为张量,这样能完整的表示高维数据并保持其本征结构,较好的去除图像的空间冗余和光谱冗余,但稀疏表示不足,达不到更高的压缩比。本算法的提出可以很好的克服以上两种方法的不足,在保证较高压缩性能的同时,有效地保护了光谱信息。首先将多光谱图像的所有谱段进行KLT,消除多光谱图像的光谱冗余。然后将变换后的每个谱段进行2维离散9/7小波变换,消除多光谱图像的空间冗余。其次,将变换后的每个小波子带都看作非负张量,对其进行Tucker分解(Tucker Decomposition, TD),并用阻尼高斯-牛顿算法(damped Gauss-Newton, dGN)求出最优解,进一步消除光谱冗余和空间残余冗余。最后,将得到的模式矩阵和核心张量进行熵编码。在压缩比4：1-32：1范围内,平均PSNR高于43dB,与其他多光谱图像压缩算法相比,极大的提高了系统PSNR,提升了多光谱图像压缩算法的性能。4、采用MT9V032型CMOS数字图像传感器设计了一款完整的小型化、低功耗相机。基于初级像差理论设计了焦距为12.95 mm,F数为5的光学系统,该系统体积小、结构紧凑,在空间频率83 lp/mm处,各视场调制传递函数(MTF)均优于0.5；电子学系统以FPGA作为时序控制平台,控制CMOS输出数字视频信号,数字视频信号通过差分芯片以低压差分信号(LVDS)格式输出到图像采集卡,最后在计算机上成像。实验结果表明,本文设计的相机像质良好、功耗低、移植性强、可靠性高,时钟为26.6 MHz时,帧频为60帧/秒,并可通过调节内部寄存器的值实现多种模式,特别适用于对相机体积以及成像质量要求较高的场合。5、根据高分辨率空间CMOS相机视频压缩系统指标,设计了用于大、中型视频压缩系统的基于KAC-06040的CMOS相机系统。设计的焦距为1175mm,F数为6.71的光学系统,该系统体积小、结构紧凑,在空间频率106.41 lp/mm处,各视场MTF均优于0.446(有遮拦),能量集中度11um以内的能量集中度均优于80%;电子学系统以FPGA作为时序控制平台,控制CMOS输出数字视频信号,数字视频信号通过Camera Link传输线以Medium模式输出到图像采集卡,最后在计算机上成像。实验结果表明,本文设计的相机像质良好、功耗低、移植性强、可靠性高,满足项目需求。
[Abstract]:Because of the continuous improvement of the spatial camera in the video resolution and frame rate, the amount of video data in the CMOS output has increased dramatically. The operation process of the existing compression algorithm is very complex, and it is mainly in the software simulation stage. The hardware implementation is difficult and the effect is poor. It is effective to alleviate on the satellite memory and channel transmission. In order to develop a real-time video compression system with good compression performance and fast data processing, the paper studies the video compression technology of all color and multispectral CMOS cameras in the background of the project "space CMOS camera technology" involved in the project. The main contents and results of this paper are summarized as follows: 1, the root of this paper is as follows According to the requirements of the project and the features of CMOS video, the ADV212 compression technology of the small CMOS camera and the high resolution CMOS camera ASIC is discussed. It is proposed to use the Custom-specific working mode to provide the video interface for various formats, and to cache the data through the table operation of the block RAM in FPGA and the DDR3 SDRAM in the FPGA, which significantly improves the work efficiency. In order to apply to different applications, in order to apply to different applications, this method realizes the transfer of the memory after the storage and the transfer between direct transmission, and improves the error correction ability of the flash memory greatly by error correction coding. Finally, in order to verify the feasibility of the method, this paper is based on the compression plate and the decompression board. The system can achieve real-time and stable work. Through the software setting, the system can achieve high compression ratio. When the compression ratio 80:1, the average peak signal to noise ratio (PSNR) is higher than 28 dB. When the compression ratio is 150:1, the average PSNR is higher than 26 dB. It solves the problem of hard parts of the compression system and the difficulty of real-time transmission under the large amount of data. A spatial multi spectral video compression algorithm is proposed by Discrete Wavelet Transform (DWT) and Karhunen-Loeve transform (KLT). Through the effective combination of two kinds of transform, the energy of the image can be concentrated to a few coefficients, and the compression effect is better. First, each spectrum segment of the multi spectral image is fast 2 dimensional discrete. 5/3 wavelet transform eliminates most of the spatial redundancy of multispectral images. Then, the wavelet coefficients generated by all spectral segments are modified to eliminate the spectral redundancy and residual spatial redundancy. Finally, the entropy coding of the obtained coefficients is used to obtain the compressed code stream. The experimental results show that the average PSNR reaches 41D within the range of 0.25-2bit/pixel (b/p). B, compared with other multi spectral image compression algorithms, it greatly improves the system PSNR and improves the performance of the multi spectral image compression algorithm. At the same time, the hardware implementation strategy is proposed to verify the correctness of the theory and the feasibility of the algorithm. It provides a reference.3 for the realization of the spatial multi spectral image compression system, the image is KLT and the first level D. After WT, there is still a great correlation between the coefficients of the different spectral segments and the high frequency subbands of the same spectral section. The average value is greater than 0.9. The value will be reduced accordingly, but the effect is not very ideal. The compression algorithm based on DWT and Tucker decomposes the image as a tensor, so that the high dimension can be expressed in a complete dimension. The data and its eigenstructure are maintained to better remove the spatial redundancy and spectral redundancy of the image, but the sparse representation is insufficient and can not reach the higher compression ratio. The proposed algorithm can overcome the shortcomings of the above two methods, and effectively protect the spectral information while ensuring high compression performance. The spectral section is KLT to eliminate the spectral redundancy of the multispectral image. Then each spectral segment after the transformation is divided into 2 dimensional discrete 9/7 wavelet transform to eliminate the spatial redundancy of the multispectral image. Secondly, each wavelet subband after transformation is regarded as a non negative tensor and Tucker decomposition (Tucker Decomposition, TD) is applied to it, and the damping Gauss Newton calculation is used. The optimal solution is obtained by the method (damped Gauss-Newton, dGN) to further eliminate spectral redundancy and space residual redundancy. Finally, the obtained pattern matrix and core tensor are entropy coded. The average PSNR is higher than 43dB in the range of compression ratio 4:1-32:1. Compared with other multispectral image compression algorithms, the system PSNR is greatly improved and the multi light is enhanced. The performance of the spectral image compression algorithm is.4. A complete miniaturized and low power consumption camera is designed by using MT9V032 CMOS digital image sensor. Based on the primary aberration theory, the optical system with a focal length of 12.95 mm and a F number of 5 is designed. The system is small in size and compact in structure. The field modulation transfer function (MTF) is superior to all field modulation transfer functions (MTF) at the space frequency rate of 83 lp/mm. 0.5, the electronic system uses FPGA as the timing control platform to control the CMOS output digital video signal. The digital video signal is output to the image acquisition card in the low voltage differential signal (LVDS) format through the differential chip, and finally is imaging on the computer. The experimental results show that the camera designed in this paper has good image quality, low power consumption, strong portability and high reliability. When the clock is 26.6 MHz, the frame rate is 60 frames per second, and many modes can be realized by adjusting the value of the internal registers. It is especially suitable for the situation of.5 with high camera volume and high imaging quality. According to the index of the video compression system of the high resolution space CMOS camera, a KAC-06040 based CMOS for large and medium video compression system is designed. Camera system. The optical system with a focal length of 1175mm and a F number of 6.71 is designed. The system is small in size and compact in structure. At the space frequency of 106.41 lp/mm, the MTF of each field of view is better than 0.446. The energy concentration within 11um of the energy concentration is better than 80%; the electronic system uses FPGA as the timing control platform to control the CMOS output digital video. The signal, the digital video signal is output to the image acquisition card by the Medium mode through the Camera Link transmission line, and finally is imaging on the computer. The experimental results show that the camera designed in this paper has good quality, low power consumption, strong portability and high reliability, and meets the requirements of the project.

【学位授予单位】：中国科学院研究生院(长春光学精密机械与物理研究所)
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】